gg

Add IMAGE_BASE_URL to the deploy env so the backend builds image URLs
against the Cloudflare image-proxy Worker instead of mackolik, and check
in the Worker source (already live on files.iddaai.com).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.gitea/workflows/deploy.yml

@@ -46,6 +46,7 @@ jobs:
             -e AI_ENGINE_URL='http://iddaai-ai-engine:8000' \
             -e JWT_SECRET='${{ secrets.JWT_SECRET }}' \
             -e JWT_ACCESS_EXPIRATION='1d' \
+            -e IMAGE_BASE_URL='https://files.iddaai.com' \
             iddaai-be:latest /bin/sh -c "npx prisma migrate deploy && node dist/src/main.js"
 
       - name: Saglik Kontrolu

ai-engine/BETTING_WORKFLOW.md

@@ -0,0 +1,313 @@
+# IDDAAI — Bahis Motoru Operasyon Workflow'u (V31d)
+
+> Bu doküman, AI bahis tahmin motorunun **nasıl çalıştırılacağı, doğrulanacağı,
+> izleneceği ve yeniden ayarlanacağına** dair operasyon kılavuzudur.
+> Hedef: **hem hacim hem kâr** — gerçekçi beklenti **premium tier'da +%30 ROI**,
+> daha geniş ağda +%5–15.
+>
+> Son güncelleme: 2026-05-29 · Judge sürümü: `judge-v31d-evidence-tiers`
+>
+> **V31d ne değiştirdi (hacim krizi çözümü):** V31c yalnızca **28 oynanabilir
+> bahis / 10k maç** üretiyordu çünkü iki veto (`calibrated_confidence_too_low`,
+> `play_score_too_low`) HER underdog'u reddediyordu — bunlar ">%45 model güveni
+> iste" diyen FAVORİ-seçme kuralı. Ama kârlı bir 6.5 oran underdog'u zaten sadece
+> ~%20 tutar; kâr oran priminden gelir. V31d, **MS değer-tier eşleşmelerinde** bu
+> iki vetoyu kaldırır ve skoru tier kalitesinden üretir. Sonuç (60g doğrulama):
+> **28 → 602 oynanabilir bahis (22x), −1.6u → +39.4u, ROI −%28 → +%32.7.**
+> Tüm zengin analiz çıktısı (market_board, v25/v27, triple_value, olasılıklar)
+> **aynen korunur** — yalnızca `playable` bayrağı değişir.
+
+---
+
+## 0. TL;DR — En Önemli 5 Kural
+
+1. **SADECE TEKLİ BAHİS OYNA. KOMBİNE YOK.** Matematiksel olarak kanıtlandı:
+   1-leg `+%3.4` → 2-leg `-%32` → 3-leg `-%67` → 4-leg `-%83`. Marjinal +EV bacakları
+   çarpmak kazancı yok eder.
+2. **Asıl kâr MS (1X2) underdog bölgesinde.** Oran ≥ 6.0 + model_gap ≥ 0 = en yüksek ROI.
+3. **Hiçbir market mute edilmez.** Tier sistemi filtreler; gerçek ROI'ler görünür kalır
+   (`MUTED_MARKETS = set()`).
+4. **Kalibrasyon ≠ Bahis sinyali.** MS tier'ları ham model olasılığını kullanır
+   (`model_gap`, `ev_edge`). İzotonik kalibratörler sadece ekrandaki `calibrated_confidence`'i
+   etkiler (BTTS/OU25'te şişik — dikkat).
+5. **Backtest'e körü körüne güvenme.** Model eğitim kesim tarihini bil; in-sample/out-of-sample
+   ayrımını her zaman yap (bkz. Bölüm 6).
+
+---
+
+## 1. Sistem Mimarisi (Pipeline)
+
+```
+Maç verisi (DB: matches, odds, elo, form, h2h…)
+        │
+        ▼
+[V25 Ensemble]  XGBoost + LightGBM + CatBoost  → her market için ham olasılık
+        │
+        ▼
+[V27 Dual-Engine]  ikinci görüş / consensus (AGREE / DISAGREE)
+        │
+        ▼
+[İzotonik Kalibrasyon]  ham olasılık → calibrated_confidence (ekran için)
+   └─ kalibratörü OLMAYAN marketlerde hafif damping (×0.92)
+        │
+        ▼
+[BettingBrain V31d — Deterministik Hâkim]
+   ├─ ev_edge      = calibrated_probability × oran − 1     (ham-prob + market blend)
+   ├─ model_gap    = ham_model_olasılık − implied_prob
+   ├─ trap_market  = market geçmiş banttan fazla fiyatlamış mı?
+   ├─ odds_reliability = lig bazında geçmiş Brier skorundan
+   └─ MARKET_ODDS_TIERS → value_tier (premium/strong/standard) → bet_grade (A/B/C)
+        │
+        ▼
+[Çıktı]  bet_summary[] → playable, value_tier, stake_units, bet_grade
+        → BE (smart-coupon) → FE / Mobile
+```
+
+**Anahtar dosyalar:**
+- `services/betting_brain.py` — deterministik hâkim, tier tanımları (`MARKET_ODDS_TIERS`)
+- `services/orchestrator/market_board.py` — ev_edge/model_gap/kalibrasyon hesapları
+- `scripts/diagnostic_backtest_multi.py` — çok-pick backtest (maç başına TÜM marketler)
+- `models/v25/`, `models/calibration/` — model ve kalibratör dosyaları
+
+---
+
+## 2. V31d — Kanıta Dayalı Kademeli Değer Sistemi (Evidence-Based Tiers)
+
+Kullanıcı risk iştahına göre seçer. Her tier maç başına ayrı sinyal üretir.
+**Sadece premium otomatik STAKE'lenir (BET); strong/standard WATCH** olarak görünür
+(tam analiz gösterilir, oynanmaz) çünkü 60 günlük veri o bantların ~başabaş olduğunu
+söylüyor.
+
+| Tier | Grade | Oran bandı | Filtre | 60g ROI* | Aksiyon | Karakter |
+|------|:----:|-----------|--------|:----:|:----:|----------|
+| **premium**  | A | **6.00 – 7.50** | model_gap ≥ 0, rel ≥ 0.30 | **+%32.7** | **BET** | Doğrulanmış edge; ~%20 hit, yüksek varyans |
+| **strong**   | B | 5.00 – 6.00 | model_gap ≥ 0, rel ≥ 0.30 | ~%−1 (başabaş) | WATCH | Görünür, oynanmaz (kanıt yetersiz) |
+| **standard** | C | 3.00 – 5.00 | model_gap ≥ 0, rel ≥ 0.30 | +%0.5 (başabaş) | WATCH | Hacim bölgesi, marj yok |
+| info (—) | — | market’e özel | ultrastrict (min_edge≥0.02, rel≥0.45-0.55, trap yok) | ~0 | REJECT/info | Bilgi amaçlı, nadiren geçer |
+
+\* 60 günlük doğrulamadan (72.582 settled satır, 7.793 maç, 2026-04-17..05-28;
+`ms_envelope.py` + `new_gate_sim.py`). premium: 602 bahis, +%32.7 ROI, +39.4u,
+%20.6 hit, **6 haftanın 6'sı da pozitif**, OOS(>05-24) +%47.4.
+
+**NEDEN 6.0–7.5 (V31c'deki 6.0–50.0 değil):** edge dar bir banda yoğunlaşmış.
+`6.0–7.0 +%35` · `7.0–8.0 ~başabaş` · **`8.0+ NEGATİF`** (−%10..−26, longshot mezarlığı).
+Eski geniş premium tier kaybeden longshot'ları içeri alıyordu. 7.5 üstünde modelin
+edge'i buharlaşıyor.
+
+**Tasarım mantığı:** premium = ROI **ve** hacim motoru (60g'de ~14 bahis/gün = bol hacim).
+Bahisçi:
+- **Düşük risk / yüksek kalite** istiyorsa → sadece **premium (A)** oyna (varsayılan).
+- **Daha fazla hacim** istiyorsa → premium bandını 6.0–8.0'e genişlet (ROI +%32.7 → +%19,
+  hâlâ sağlam, +%44 hacim) — `MARKET_ODDS_TIERS["MS"]` premium `max_odds`'u değiştir.
+
+**Non-MS marketler (DC, OU25, OU35, BTTS, HT, OU15, HTFT, OE, HT_OU05, HT_OU15, CARDS):**
+hepsi `ultrastrict` tek-tier ile bilgi amaçlı. Geçmiş veride sistematik olarak kayıp
+verdikleri için BET üretmeleri zorlaştırıldı (mute YOK — sadece sıkı eşik).
+
+**Veto mantığı (V31d kritik):** value-tier eşleşmelerinde `calibrated_confidence_too_low`
+ve `play_score_too_low` vetoları KALDIRILIR (bunlar favori-seçme kuralı). Ama gerçek
+koruma vetoları AKTİF kalır: `extreme_negative_ev` (ev<−0.20), `ev_edge_too_high_trap`
+(ev≥0.30), `htft_reversal_risk_high`, `v25_v27_hard_disagreement`, `low_reliability_hard`.
+60g'de premium tier-eşleşmelerinin ~%71'i oynanabilir oldu; kalan ~%29 bu koruma
+vetolarıyla doğru şekilde reddedildi.
+
+---
+
+## 3. EN İYİ BAHİS DEĞERLERİ — Kesin Sıralama (Best Bet Values)
+
+> "Multi bahislerde bütün bahis değerlerinin en iyisi" sorusunun cevabı.
+> **Hepsi TEKLİ oynanır.** (Aşağıdaki ROI'ler 0.2u sabit stake simülasyonundan.)
+
+### MS (1X2) underdog — ince oran-bandı haritası (60g, gap ≥ 0)
+
+> "Hangi bahis hangi oranda tutuyor" sorusunun kesin cevabı. `ms_envelope.py`.
+> drop-3/5 = en büyük 3/5 kazancı çıkarınca ROI (konsantrasyon/sağlamlık testi).
+
+| Oran bandı | Bahis | Hit% | ROI | drop-3 ROI | Karar |
+|-----------|------:|-----:|----:|-----:|:-----:|
+| **6.0 – 6.5** | 469 | %22.0 | **+%37.7** | +%34.4 | ✅ elit |
+| **6.0 – 7.0** | 492 | %21.5 | **+%35.2** | +%29.9 | ✅ elit, sağlam |
+| **6.0 – 7.5** (premium) | 645 | %20.0 | **+%29.3** | +%24.4 | ✅ ÖNERİLEN |
+| 6.0 – 8.0 | 928 | %17.7 | +%19.1 | +%15.5 | ✅ hacim opsiyonu |
+| 7.5 – 8.0 | 283 | %12.4 | −%4.0 | — | ❌ |
+| 8.0 – 9.0 | 78 | %9.0 | −%25.7 | — | ❌ longshot |
+| 9.0+ | ~266 | <%10 | negatif | — | ❌ mezarlık |
+| 5.0 – 6.0 (strong) | ~1000 | %18 | ~−%1 | — | ⚠️ başabaş → WATCH |
+| 3.0 – 5.0 (standard) | ~5745 | %27 | +%0.5 | — | ⚠️ başabaş → WATCH |
+
+**Korumalı premium (htft/disagreement vetoları uygulanmış) = staked set:**
+602 bahis · %20.6 hit · **+%32.7 ROI** · +39.4u · 6/6 hafta pozitif · OOS +%47.4.
+
+**Okuma:** Edge tamamen **6.0–7.5** bandında. 8.0 üstü longshot'lar kaybeder
+(eski 6.0–50.0 premium tier'ı bu yüzden sulandırıyordu). 5.0 altı başabaş.
+Premium tek başına ~14 bahis/gün = hem hacim hem +%32.7 ROI.
+
+### ❌ İşe YARAMAYAN yapılandırmalar
+- **Kombine (parlay):** her ek bacak ROI'yi çökertir (yukarıdaki TL;DR).
+- **MS 8.0+ longshot:** −%10..−26 ROI, model edge'i yok.
+- **MS 5.0–6.0 / 3.0–5.0:** başabaş; WATCH olarak göster, stake'leme.
+- **OU25 her konfigürasyon:** sistematik kayıp (60g'de OU25 −%22.8, OU35 −%17.2).
+- **BTTS:** sadece çok yüksek reliability'de marjinal.
+
+---
+
+## 4. KRİTİK KURAL — Tekli Bahis, Kombine Yok
+
+| Kupon tipi | Hit% | ROI | Sonuç |
+|-----------|-----:|----:|:-----:|
+| 1-leg (tekli) | ~%24 | **+%3.4** | ✅ |
+| 2-leg | düşük | −%32.4 | ❌ |
+| 3-leg | çok düşük | −%66.6 | ❌ |
+| 4-leg | minimal | −%83.0 | ❌ |
+
+**Neden:** Tekil bacaklar yalnızca marjinal +EV. Kombine, kazanma olasılıklarını
+çarparken (her biri <1) kayıp olasılığını üssel büyütür. Düz (flat) tekli stake
+matematiksel olarak üstündür. **Ürün, kullanıcıyı kombineye teşvik etmemeli;**
+"günün premium tekli değerleri" şeklinde sunmalı.
+
+---
+
+## 5. Önerilen Stake Politikası
+
+- **Flat stake** (sabit birim) — Kelly değil. Marjinal edge'de Kelly varyansı patlatır.
+- **premium (A): 0.5u sabit** (`VALUE_TIER_STAKE_UNITS`). ~%20 hit + uzun kayıp serileri
+  (60g'de en uzun 35 ardışık kayıp) nedeniyle KÜÇÜK tutulur — kâr **frekanstan** gelir,
+  bahis başı büyüklükten değil. Bankroll/risk iştahı izin veriyorsa artırılabilir.
+- strong/standard WATCH = stake YOK (görünür ama oynanmaz).
+- Günlük/maç başına 1 sinyal; aynı maça birden çok tier'dan bahis = korelasyon riski,
+  en yüksek value_tier'ı seç.
+- **Drawdown uyarısı:** 0.5u'da en kötü tarihsel düşüş ≈ −34u; 35 ardışık kayıp mümkün.
+  Bu bir maraton stratejisidir — kısa vadeli sonuçlara göre stake değiştirme.
+
+---
+
+## 6. Backtest Metodolojisi & Leakage Disiplini ⚠️
+
+**En kritik bölüm. Backtest sayıları yanlış yorumlanırsa sistem kârlı sanılıp kaybettirir.**
+
+### 6.1 Komut
+```bash
+# Konteyner içinde:
+python scripts/diagnostic_backtest_multi.py --days 60 --max-matches 10000 \
+    --progress-interval 100 --checkpoint-every 200
+# Çıktı: reports/multi_backtest_YYYYMMDD.{csv,json,txt}
+# Checkpoint'li → kesilirse kaldığı yerden devam eder.
+```
+
+### 6.2 Lookahead / Sızıntı (leakage) kontrolü — ZORUNLU
+- **Feature lookahead:** ✅ temiz — feature'lar match_date ÖNCESİ veriden hesaplanıyor.
+- **Model eğitim-seti üyeliği:** Bunu HER ZAMAN kontrol et. Kalibratörler
+  `models/calibration/*_metrics.json` içindeki `last_trained` tarihinde, son ~5000
+  maç üzerinde fit edilir. Backtest penceresi bu tarihle çakışırsa **calibrated_confidence
+  in-sample (şişik)** olur.
+- **Pratik test (ucuz):** Backtest sonucunu eğitim kesim tarihine göre ikiye böl;
+  in-sample vs out-of-sample hit% karşılaştır. Tüm-market hit% **neredeyse aynıysa**
+  (örn. %49.7 vs %49.4) → temel modellerde anlamlı sızıntı YOK, edge gerçek.
+  Eski veride hit% **aniden yükseliyorsa** → o dönem eğitim setinde, ROI'yi yok say.
+  - Hazır script: `/tmp/leakage_split.py <csv>` (eğitim tarihine göre böler).
+- **Geriye doğru ne kadar gidilebilir?** Modeller en son holdout penceresini (≈son
+  10k maç ≈ 60-70 gün) eğitimden hariç tutuyor. Bu yüzden **~60 gün geriye backtest
+  çoğunlukla temiz holdout'tur.** Daha geriye (90+ gün) gitmek eğitim setine girip
+  ROI'yi yapay iyi gösterebilir → kaçın.
+
+### 6.3 Doğrulama scriptleri
+- `/tmp/v31c_validation.py <csv>` — V31c tier dökümü (premium/strong/standard ROI).
+- `/tmp/best_bet_values.py <csv>` — grid-search liderlik tablosu + portföy + kombine testi.
+- `/tmp/leakage_split.py <csv>` — in/out-of-sample sızıntı probu.
+
+### 6.4 Doğrulama eşiği (bir tier "kârlı" sayılmadan önce)
+- n ≥ 50 bahis (tercihen ≥ 200), out-of-sample.
+- ROI > 0 hem in- hem out-of-sample'da, ya da en azından OOS'ta çökmemiş.
+- Kümülatif kâr eğrisi yukarı trend (tek bir şanslı güne bağlı değil).
+
+---
+
+## 7. Operasyonel Döngü (Cadence)
+
+### Günlük
+- Motor sağlık kontrolü (futbol pipeline çalışıyor mu; basketbol `readiness_summary`
+  hatası bilinen/zararsız).
+- Günün sinyallerini üret; **premium (A) tekli** değerleri öne çıkar.
+- Settle olan dünün bahislerini logla (gerçek hit/ROI takibi).
+
+### Haftalık
+- Son 7-14 günün gerçek sonuçlarını backtest tahminiyle karşılaştır (calibration drift).
+- Tier bazında gerçekleşen ROI'yi izle; standard (C) sürekli negatifse eşik sıkılaştır.
+
+### Aylık
+- Modelleri yeniden eğit (Colab: `extract_training_data_v27.py` → eğitim → `fetch_xgb_models.sh`).
+- **Yeniden eğitimden sonra MUTLAKA** 60 günlük backtest + leakage_split ile yeniden doğrula.
+- Tier eşiklerini güncelle (Bölüm 8).
+- `models/calibration/*_metrics.json` `last_trained` tarihini not et (bir sonraki
+  backtest'in OOS penceresini bilmek için).
+
+---
+
+## 8. Tier / Eşik Güncelleme Protokolü
+
+1. Yeni backtest CSV'sini al → `v31c_validation.py` + `leakage_split.py` çalıştır.
+2. Her tier için OOS ROI'ye bak:
+   - ROI sağlam pozitif + n yeterli → koru.
+   - ROI marjinal/negatif → oran bandını daralt veya min_reliability/min_model_gap yükselt.
+   - premium 6.0+ eşiği: OOS'ta hâlâ en iyi ROI mi? Değilse bandı kaydır (örn. 6.5+).
+3. `betting_brain.py` → `MARKET_ODDS_TIERS` düzenle, **versiyon string'ini artır**
+   (`judge-v31c-…` → `judge-v31d-…`).
+4. Lokal syntax kontrol → sunucuya deploy (Bölüm 9) → yeniden doğrula.
+5. Tier'lar netleştikten SONRA `value_tier`'ı UI'a yay (BE smart-coupon → FE badge → mobil).
+
+---
+
+## 9. Deploy Prosedürü (AI Engine)
+
+```bash
+# 1. Lokal syntax kontrol
+python3 -c "import ast; ast.parse(open('services/betting_brain.py').read())"
+
+# 2. Sunucuya kopyala  (SSH: port 2222, kullanıcı haruncan)
+scp -P 2222 services/betting_brain.py haruncan@<host>:/tmp/betting_brain.py
+
+# 3. Konteynere koy + import testi
+docker cp /tmp/betting_brain.py iddaai-ai-engine:/app/services/betting_brain.py
+docker exec iddaai-ai-engine python -c "from services.betting_brain import BettingBrain; print('OK')"
+
+# 4. Yeniden başlat + doğrula
+docker restart iddaai-ai-engine
+docker exec iddaai-ai-engine python -c "from services.betting_brain import BettingBrain as B; \
+  print([t['value_tier'] for t in B().MARKET_ODDS_TIERS['MS']])"
+```
+> Not: Port 8000 host-localhost'a expose DEĞİL; sağlık testini konteyner içinden veya
+> Docker network üzerinden yap. Basketbol `readiness_summary` hatası bilinen, bloklamıyor.
+
+---
+
+## 10. Bilinen Sınırlamalar & Uyarılar
+
+- **Kalibrasyon şişmesi:** BTTS / OU25 izotonik kalibratörleri olasılığı %10-15 fazla
+  gösteriyor (overcalibrated). Bu marketlerde ekrandaki `calibrated_confidence`'e tam
+  güvenme; bahis kararı zaten ham-prob `model_gap`/`ev_edge` ile veriliyor.
+- **Out-of-sample örneklem küçük:** Eğitim kesim tarihinden sonraki temiz pencere dar
+  olabilir (~200 MS bahsi). İstatistiksel kesinlik için ileriye doğru gerçek sonuç
+  biriktir (paper-trade) veya 60 günlük holdout backtest kullan.
+- **standard (C) tier kırılgan:** in-sample +%0.4, küçük OOS örnekte negatife düşebiliyor.
+  Hacim için var; ROI garantisi değil.
+- **Tek pencere overfit riski:** Tek bir sezon/dönem penceresine göre ayar yapma;
+  farklı lig/sezon çeşitliliği ara.
+- **Basketbol:** `BasketballV25Predictor.readiness_summary` eksik — futbolu etkilemiyor,
+  ayrı düzeltilecek.
+
+---
+
+## 11. Hızlı Komut Referansı
+
+```bash
+# 60 günlük backtest (konteyner içi)
+python scripts/diagnostic_backtest_multi.py --days 60 --max-matches 10000
+
+# Doğrulama (CSV lokale çekildikten sonra)
+python3 /tmp/v31c_validation.py  reports/multi_backtest_YYYYMMDD.csv
+python3 /tmp/best_bet_values.py  reports/multi_backtest_YYYYMMDD.csv
+python3 /tmp/leakage_split.py    reports/multi_backtest_YYYYMMDD.csv
+
+# Kalibratör eğitim tarihleri
+grep -o '"last_trained":[^,]*' models/calibration/*.json
+```

ai-engine/catboost_info/catboost_training.json

@@ -1,874 +0,0 @@
-{
-"meta":{"test_sets":["test"],"test_metrics":[{"best_value":"Min","name":"Logloss"}],"learn_metrics":[{"best_value":"Min","name":"Logloss"}],"launch_mode":"Train","parameters":"","iteration_count":2000,"learn_sets":["learn"],"name":"experiment"},
-"iterations":[
-{"learn":[0.692389481],"iteration":0,"passed_time":0.04679785798,"remaining_time":93.54891809,"test":[0.6924099937]},
-{"learn":[0.6916338586],"iteration":1,"passed_time":0.08350330552,"remaining_time":83.41980222,"test":[0.6916660956]},
-{"learn":[0.6910159214],"iteration":2,"passed_time":0.132821758,"remaining_time":88.41501689,"test":[0.691108145]},
-{"learn":[0.6903417151],"iteration":3,"passed_time":0.162826233,"remaining_time":81.25029026,"test":[0.6904585078]},
-{"learn":[0.6896961461],"iteration":4,"passed_time":0.1969265393,"remaining_time":78.57368918,"test":[0.689812816]},
-{"learn":[0.6890979366],"iteration":5,"passed_time":0.2309352918,"remaining_time":76.74749531,"test":[0.689192261]},
-{"learn":[0.6884946167],"iteration":6,"passed_time":0.2693987513,"remaining_time":76.70167304,"test":[0.6886032715]},
-{"learn":[0.6879503686],"iteration":7,"passed_time":0.3199759681,"remaining_time":79.67401607,"test":[0.6880706742]},
-{"learn":[0.6874528094],"iteration":8,"passed_time":0.3645802206,"remaining_time":80.65324659,"test":[0.6876192378]},
-{"learn":[0.6869036785],"iteration":9,"passed_time":0.4116507506,"remaining_time":81.91849936,"test":[0.6870868859]},
-{"learn":[0.6863761921],"iteration":10,"passed_time":0.4562469316,"remaining_time":82.49774064,"test":[0.6865493528]},
-{"learn":[0.6859038678],"iteration":11,"passed_time":0.491541699,"remaining_time":81.43207481,"test":[0.686105086]},
-{"learn":[0.685410175],"iteration":12,"passed_time":0.5221556769,"remaining_time":79.80948692,"test":[0.6856345086]},
-{"learn":[0.6849483392],"iteration":13,"passed_time":0.5553110353,"remaining_time":78.77483686,"test":[0.6852027185]},
-{"learn":[0.6845417792],"iteration":14,"passed_time":0.5952927147,"remaining_time":78.77706925,"test":[0.6848238481]},
-{"learn":[0.6841038875],"iteration":15,"passed_time":0.6300274185,"remaining_time":78.12339989,"test":[0.6844045699]},
-{"learn":[0.6836957422],"iteration":16,"passed_time":0.662600544,"remaining_time":77.29040464,"test":[0.6840077621]},
-{"learn":[0.6832947461],"iteration":17,"passed_time":0.7004221698,"remaining_time":77.12426337,"test":[0.6836197496]},
-{"learn":[0.6829014105],"iteration":18,"passed_time":0.7300844347,"remaining_time":76.12090869,"test":[0.6832475033]},
-{"learn":[0.6825264546],"iteration":19,"passed_time":0.7641559459,"remaining_time":75.65143865,"test":[0.6829012069]},
-{"learn":[0.6822106577],"iteration":20,"passed_time":0.8040792063,"remaining_time":75.77489282,"test":[0.6825880966]},
-{"learn":[0.6818649349],"iteration":21,"passed_time":0.8356039756,"remaining_time":75.12839381,"test":[0.6822424968]},
-{"learn":[0.6815467855],"iteration":22,"passed_time":0.8861440327,"remaining_time":76.16985881,"test":[0.6819180513]},
-{"learn":[0.6812293319],"iteration":23,"passed_time":0.920219319,"remaining_time":75.76472393,"test":[0.6816384467]},
-{"learn":[0.6808837443],"iteration":24,"passed_time":0.960164738,"remaining_time":75.8530143,"test":[0.6813262593]},
-{"learn":[0.6805816494],"iteration":25,"passed_time":0.9895547925,"remaining_time":75.13004463,"test":[0.6810353411]},
-{"learn":[0.6803209634],"iteration":26,"passed_time":1.025550161,"remaining_time":74.94112844,"test":[0.6808138172]},
-{"learn":[0.6800350862],"iteration":27,"passed_time":1.060852064,"remaining_time":74.71429535,"test":[0.6805550049]},
-{"learn":[0.6797703947],"iteration":28,"passed_time":1.10467538,"remaining_time":75.07983357,"test":[0.680347991]},
-{"learn":[0.6794926675],"iteration":29,"passed_time":1.141766834,"remaining_time":74.97602208,"test":[0.680089679]},
-{"learn":[0.6792251865],"iteration":30,"passed_time":1.180421588,"remaining_time":74.9758099,"test":[0.6798451919]},
-{"learn":[0.6789670166],"iteration":31,"passed_time":1.213674604,"remaining_time":74.64098814,"test":[0.6796090443]},
-{"learn":[0.678722402],"iteration":32,"passed_time":1.245848393,"remaining_time":74.26011482,"test":[0.6793890865]},
-{"learn":[0.678476935],"iteration":33,"passed_time":1.287262512,"remaining_time":74.43406171,"test":[0.6791683772]},
-{"learn":[0.6782297335],"iteration":34,"passed_time":1.327473991,"remaining_time":74.52818262,"test":[0.6789766369]},
-{"learn":[0.6780226701],"iteration":35,"passed_time":1.3760549,"remaining_time":75.07143955,"test":[0.6787930242]},
-{"learn":[0.6778291026],"iteration":36,"passed_time":1.427620019,"remaining_time":75.74102965,"test":[0.6786087714]},
-{"learn":[0.6776045324],"iteration":37,"passed_time":1.468182407,"remaining_time":75.80457587,"test":[0.6784161299]},
-{"learn":[0.6773969079],"iteration":38,"passed_time":1.508647379,"remaining_time":75.85788487,"test":[0.6782227897]},
-{"learn":[0.6771819602],"iteration":39,"passed_time":1.549435187,"remaining_time":75.92232419,"test":[0.6780242369]},
-{"learn":[0.6769816736],"iteration":40,"passed_time":1.586036608,"remaining_time":75.78160282,"test":[0.6778499631]},
-{"learn":[0.6767984027],"iteration":41,"passed_time":1.621458864,"remaining_time":75.59086802,"test":[0.6776975784]},
-{"learn":[0.6766201184],"iteration":42,"passed_time":1.663424818,"remaining_time":75.70517136,"test":[0.6775231674]},
-{"learn":[0.6764394377],"iteration":43,"passed_time":1.70110089,"remaining_time":75.62166686,"test":[0.6773582124]},
-{"learn":[0.6762698797],"iteration":44,"passed_time":1.739954496,"remaining_time":75.59135644,"test":[0.6772234666]},
-{"learn":[0.6760974263],"iteration":45,"passed_time":1.776461223,"remaining_time":75.46098325,"test":[0.6770659843]},
-{"learn":[0.6759245179],"iteration":46,"passed_time":1.819761638,"remaining_time":75.61690381,"test":[0.6769049529]},
-{"learn":[0.6757673909],"iteration":47,"passed_time":1.869479807,"remaining_time":76.02551217,"test":[0.6767664194]},
-{"learn":[0.6756172628],"iteration":48,"passed_time":1.916010121,"remaining_time":76.28848462,"test":[0.6766584917]},
-{"learn":[0.675474531],"iteration":49,"passed_time":1.953635244,"remaining_time":76.19177452,"test":[0.6765507257]},
-{"learn":[0.6753286933],"iteration":50,"passed_time":1.993876686,"remaining_time":76.19736591,"test":[0.6764489911]},
-{"learn":[0.6751900513],"iteration":51,"passed_time":2.038943041,"remaining_time":76.38194316,"test":[0.6763947956]},
-{"learn":[0.6750574835],"iteration":52,"passed_time":2.080276765,"remaining_time":76.42073325,"test":[0.6762778712]},
-{"learn":[0.6749329567],"iteration":53,"passed_time":2.158576742,"remaining_time":77.78871001,"test":[0.6761865366]},
-{"learn":[0.6748033265],"iteration":54,"passed_time":2.220619687,"remaining_time":78.52918711,"test":[0.6760679685]},
-{"learn":[0.6746797823],"iteration":55,"passed_time":2.286959228,"remaining_time":79.39015604,"test":[0.6759774874]},
-{"learn":[0.674535525],"iteration":56,"passed_time":2.328472096,"remaining_time":79.3723032,"test":[0.6758500622]},
-{"learn":[0.6744256514],"iteration":57,"passed_time":2.367031568,"remaining_time":79.25474665,"test":[0.6757625065]},
-{"learn":[0.674310819],"iteration":58,"passed_time":2.409161286,"remaining_time":79.25732298,"test":[0.6756876412]},
-{"learn":[0.6741967947],"iteration":59,"passed_time":2.444825903,"remaining_time":79.04937087,"test":[0.6756151069]},
-{"learn":[0.6740879654],"iteration":60,"passed_time":2.48484996,"remaining_time":78.98564055,"test":[0.6755303655]},
-{"learn":[0.6739772476],"iteration":61,"passed_time":2.521603395,"remaining_time":78.8204416,"test":[0.6754565036]},
-{"learn":[0.67388281],"iteration":62,"passed_time":2.554102332,"remaining_time":78.5285114,"test":[0.6753738983]},
-{"learn":[0.6737789726],"iteration":63,"passed_time":2.593937938,"remaining_time":78.46662263,"test":[0.6752897299]},
-{"learn":[0.6736812332],"iteration":64,"passed_time":2.623889155,"remaining_time":78.11116175,"test":[0.6752115539]},
-{"learn":[0.6735930009],"iteration":65,"passed_time":2.660795108,"remaining_time":77.96935967,"test":[0.6751595431]},
-{"learn":[0.6734947116],"iteration":66,"passed_time":2.695822592,"remaining_time":77.77649358,"test":[0.6750764658]},
-{"learn":[0.6733961481],"iteration":67,"passed_time":2.725876686,"remaining_time":77.44696703,"test":[0.6750179194]},
-{"learn":[0.6732990195],"iteration":68,"passed_time":2.761848366,"remaining_time":77.29172746,"test":[0.6749408803]},
-{"learn":[0.6732133575],"iteration":69,"passed_time":2.791847449,"remaining_time":76.97522253,"test":[0.6748795802]},
-{"learn":[0.673111539],"iteration":70,"passed_time":2.824541003,"remaining_time":76.73999429,"test":[0.674790372]},
-{"learn":[0.6730080451],"iteration":71,"passed_time":2.861023716,"remaining_time":76.61185729,"test":[0.6747239773]},
-{"learn":[0.6729157861],"iteration":72,"passed_time":2.897136588,"remaining_time":76.47646857,"test":[0.6746701254]},
-{"learn":[0.6728347949],"iteration":73,"passed_time":2.935718661,"remaining_time":76.40802894,"test":[0.6746120937]},
-{"learn":[0.6727640693],"iteration":74,"passed_time":3.040023476,"remaining_time":78.02726921,"test":[0.6745550085]},
-{"learn":[0.6726808811],"iteration":75,"passed_time":3.097341794,"remaining_time":78.41165279,"test":[0.6744855074]},
-{"learn":[0.6726029645],"iteration":76,"passed_time":3.152948955,"remaining_time":78.74182909,"test":[0.6744264172]},
-{"learn":[0.6725356026],"iteration":77,"passed_time":3.216126808,"remaining_time":79.24866314,"test":[0.674381715]},
-{"learn":[0.6724606887],"iteration":78,"passed_time":3.256861302,"remaining_time":79.19532355,"test":[0.6743331681]},
-{"learn":[0.6723849561],"iteration":79,"passed_time":3.305679851,"remaining_time":79.33631641,"test":[0.67428564]},
-{"learn":[0.6723050519],"iteration":80,"passed_time":3.348083566,"remaining_time":79.32064647,"test":[0.6742202413]},
-{"learn":[0.6722508802],"iteration":81,"passed_time":3.38129387,"remaining_time":79.08928832,"test":[0.6741620971]},
-{"learn":[0.6721773904],"iteration":82,"passed_time":3.41660066,"remaining_time":78.91112609,"test":[0.6741109453]},
-{"learn":[0.6721007598],"iteration":83,"passed_time":3.48099347,"remaining_time":79.39980344,"test":[0.6740556003]},
-{"learn":[0.6720353564],"iteration":84,"passed_time":3.535359896,"remaining_time":79.64957884,"test":[0.6740146772]},
-{"learn":[0.6719790902],"iteration":85,"passed_time":3.581806996,"remaining_time":79.71603012,"test":[0.673983295]},
-{"learn":[0.6719140024],"iteration":86,"passed_time":3.612293661,"remaining_time":79.42893993,"test":[0.6739595301]},
-{"learn":[0.6718573633],"iteration":87,"passed_time":3.644530261,"remaining_time":79.18570293,"test":[0.6739336659]},
-{"learn":[0.671795602],"iteration":88,"passed_time":3.67809653,"remaining_time":78.97575809,"test":[0.673890361]},
-{"learn":[0.6717369134],"iteration":89,"passed_time":3.712417516,"remaining_time":78.78574951,"test":[0.673863586]},
-{"learn":[0.6716711079],"iteration":90,"passed_time":3.743502971,"remaining_time":78.53128759,"test":[0.6738190616]},
-{"learn":[0.6716070843],"iteration":91,"passed_time":3.775351679,"remaining_time":78.2975109,"test":[0.6737799295]},
-{"learn":[0.6715517232],"iteration":92,"passed_time":3.806186247,"remaining_time":78.04728142,"test":[0.6737364374]},
-{"learn":[0.6714957378],"iteration":93,"passed_time":3.83798807,"remaining_time":77.82133257,"test":[0.6737093719]},
-{"learn":[0.6714364567],"iteration":94,"passed_time":3.871278973,"remaining_time":77.62933099,"test":[0.6736630475]},
-{"learn":[0.6713881758],"iteration":95,"passed_time":3.913531039,"remaining_time":77.6183656,"test":[0.67364367]},
-{"learn":[0.6713336502],"iteration":96,"passed_time":3.945433866,"remaining_time":77.40371802,"test":[0.6735998081]},
-{"learn":[0.6712700267],"iteration":97,"passed_time":3.989716281,"remaining_time":77.43306496,"test":[0.6735526984]},
-{"learn":[0.6712154424],"iteration":98,"passed_time":4.020621946,"remaining_time":77.20406384,"test":[0.6735012924]},
-{"learn":[0.6711600413],"iteration":99,"passed_time":4.053732144,"remaining_time":77.02091074,"test":[0.6734818024]},
-{"learn":[0.6711060533],"iteration":100,"passed_time":4.084124711,"remaining_time":76.78963194,"test":[0.6734379341]},
-{"learn":[0.6710494943],"iteration":101,"passed_time":4.116434744,"remaining_time":76.59797199,"test":[0.6734059869]},
-{"learn":[0.6709936897],"iteration":102,"passed_time":4.148330356,"remaining_time":76.40177365,"test":[0.6733740852]},
-{"learn":[0.6709472183],"iteration":103,"passed_time":4.176511193,"remaining_time":76.14101176,"test":[0.6733330971]},
-{"learn":[0.6708914508],"iteration":104,"passed_time":4.2025065,"remaining_time":75.84523636,"test":[0.6733060254]},
-{"learn":[0.6708388195],"iteration":105,"passed_time":4.232975206,"remaining_time":75.63448151,"test":[0.6732755898]},
-{"learn":[0.6707885854],"iteration":106,"passed_time":4.261364958,"remaining_time":75.39031649,"test":[0.6732294722]},
-{"learn":[0.6707454167],"iteration":107,"passed_time":4.290824713,"remaining_time":75.1688922,"test":[0.6732035176]},
-{"learn":[0.6706973013],"iteration":108,"passed_time":4.324192493,"remaining_time":75.01878903,"test":[0.673196437]},
-{"learn":[0.6706577031],"iteration":109,"passed_time":4.351512102,"remaining_time":74.76688976,"test":[0.6731652709]},
-{"learn":[0.67061108],"iteration":110,"passed_time":4.38641502,"remaining_time":74.64808984,"test":[0.673138808]},
-{"learn":[0.6705625485],"iteration":111,"passed_time":4.424063991,"remaining_time":74.57707871,"test":[0.6731062725]},
-{"learn":[0.6705146484],"iteration":112,"passed_time":4.45863849,"remaining_time":74.45531709,"test":[0.6730726625]},
-{"learn":[0.6704704423],"iteration":113,"passed_time":4.497153675,"remaining_time":74.40027922,"test":[0.6730285927]},
-{"learn":[0.6704155922],"iteration":114,"passed_time":4.533368584,"remaining_time":74.30782417,"test":[0.6729872702]},
-{"learn":[0.6703687117],"iteration":115,"passed_time":4.564651269,"remaining_time":74.13623268,"test":[0.6729721425]},
-{"learn":[0.6703324232],"iteration":116,"passed_time":4.596824343,"remaining_time":73.98136956,"test":[0.6729564624]},
-{"learn":[0.6702884624],"iteration":117,"passed_time":4.628377967,"remaining_time":73.81870623,"test":[0.6729312424]},
-{"learn":[0.670253478],"iteration":118,"passed_time":4.668052254,"remaining_time":73.78660748,"test":[0.6729354345]},
-{"learn":[0.6702140804],"iteration":119,"passed_time":4.692108266,"remaining_time":73.50969617,"test":[0.6729085401]},
-{"learn":[0.6701682529],"iteration":120,"passed_time":4.723741667,"remaining_time":73.354633,"test":[0.6728898322]},
-{"learn":[0.6701320588],"iteration":121,"passed_time":4.756626425,"remaining_time":73.22085595,"test":[0.6728773638]},
-{"learn":[0.6700939824],"iteration":122,"passed_time":4.788008428,"remaining_time":73.06578714,"test":[0.6728618874]},
-{"learn":[0.6700655902],"iteration":123,"passed_time":4.815546648,"remaining_time":72.85456058,"test":[0.6728540413]},
-{"learn":[0.6700190743],"iteration":124,"passed_time":4.843186806,"remaining_time":72.64780209,"test":[0.6728441291]},
-{"learn":[0.6699792296],"iteration":125,"passed_time":4.875548614,"remaining_time":72.51411192,"test":[0.672815631]},
-{"learn":[0.6699379404],"iteration":126,"passed_time":4.916953662,"remaining_time":72.51538748,"test":[0.6728082021]},
-{"learn":[0.669895454],"iteration":127,"passed_time":4.952918369,"remaining_time":72.43643115,"test":[0.6727900064]},
-{"learn":[0.6698563938],"iteration":128,"passed_time":4.991585558,"remaining_time":72.39733782,"test":[0.6727649552]},
-{"learn":[0.6698215571],"iteration":129,"passed_time":5.028084166,"remaining_time":72.32705685,"test":[0.6727467657]},
-{"learn":[0.6697857067],"iteration":130,"passed_time":5.059198996,"remaining_time":72.18048033,"test":[0.6727396032]},
-{"learn":[0.6697449303],"iteration":131,"passed_time":5.096035515,"remaining_time":72.1166238,"test":[0.6727245271]},
-{"learn":[0.6697052425],"iteration":132,"passed_time":5.125282589,"remaining_time":71.94663604,"test":[0.6726955143]},
-{"learn":[0.6696695553],"iteration":133,"passed_time":5.156392608,"remaining_time":71.80469109,"test":[0.67269209]},
-{"learn":[0.6696269265],"iteration":134,"passed_time":5.190402292,"remaining_time":71.70444647,"test":[0.672677932]},
-{"learn":[0.6695969271],"iteration":135,"passed_time":5.221466142,"remaining_time":71.56480065,"test":[0.6726540285]},
-{"learn":[0.6695489786],"iteration":136,"passed_time":5.251144663,"remaining_time":71.40790151,"test":[0.6726288583]},
-{"learn":[0.6695173859],"iteration":137,"passed_time":5.274361693,"remaining_time":71.16566285,"test":[0.6725863431]},
-{"learn":[0.6694811164],"iteration":138,"passed_time":5.309398952,"remaining_time":71.08483058,"test":[0.6725837967]},
-{"learn":[0.6694477439],"iteration":139,"passed_time":5.344693175,"remaining_time":71.00806646,"test":[0.6725772977]},
-{"learn":[0.6694082161],"iteration":140,"passed_time":5.377737126,"remaining_time":70.90222211,"test":[0.6725685594]},
-{"learn":[0.6693679185],"iteration":141,"passed_time":5.416087925,"remaining_time":70.8668406,"test":[0.6725553829]},
-{"learn":[0.6693341916],"iteration":142,"passed_time":5.452286939,"remaining_time":70.80347444,"test":[0.6725484347]},
-{"learn":[0.6692933159],"iteration":143,"passed_time":5.490006789,"remaining_time":70.7600875,"test":[0.6725306172]},
-{"learn":[0.6692619696],"iteration":144,"passed_time":5.521869859,"remaining_time":70.64185233,"test":[0.672543149]},
-{"learn":[0.6692229289],"iteration":145,"passed_time":5.553520721,"remaining_time":70.5221056,"test":[0.6725196247]},
-{"learn":[0.6691840164],"iteration":146,"passed_time":5.582178524,"remaining_time":70.3658286,"test":[0.6725226452]},
-{"learn":[0.6691581406],"iteration":147,"passed_time":5.611368671,"remaining_time":70.21793769,"test":[0.6725056913]},
-{"learn":[0.6691177196],"iteration":148,"passed_time":5.636941079,"remaining_time":70.02669757,"test":[0.6724771476]},
-{"learn":[0.6690851126],"iteration":149,"passed_time":5.673704689,"remaining_time":69.97569117,"test":[0.6724439435]},
-{"learn":[0.6690518144],"iteration":150,"passed_time":5.706346207,"remaining_time":69.87439826,"test":[0.672442532]},
-{"learn":[0.6690149711],"iteration":151,"passed_time":5.738210991,"remaining_time":69.76456521,"test":[0.6724303064]},
-{"learn":[0.668993877],"iteration":152,"passed_time":5.765951318,"remaining_time":69.60596133,"test":[0.6724235788]},
-{"learn":[0.6689596579],"iteration":153,"passed_time":5.795573467,"remaining_time":69.47161442,"test":[0.6724294499]},
-{"learn":[0.6689372651],"iteration":154,"passed_time":5.81744896,"remaining_time":69.24640858,"test":[0.6724285935]},
-{"learn":[0.6689003045],"iteration":155,"passed_time":5.853529431,"remaining_time":69.19171968,"test":[0.6724172017]},
-{"learn":[0.6688680182],"iteration":156,"passed_time":5.888380392,"remaining_time":69.12283479,"test":[0.6724130745]},
-{"learn":[0.6688348164],"iteration":157,"passed_time":5.924601775,"remaining_time":69.07035741,"test":[0.6723860878]},
-{"learn":[0.6687947046],"iteration":158,"passed_time":5.964531924,"remaining_time":69.06102687,"test":[0.6723707604]},
-{"learn":[0.6687605251],"iteration":159,"passed_time":5.996805452,"remaining_time":68.9632627,"test":[0.6723566111]},
-{"learn":[0.668726253],"iteration":160,"passed_time":6.022341459,"remaining_time":68.78935368,"test":[0.6723469906]},
-{"learn":[0.6686862718],"iteration":161,"passed_time":6.05082584,"remaining_time":68.65072774,"test":[0.6723287161]},
-{"learn":[0.668663478],"iteration":162,"passed_time":6.079027554,"remaining_time":68.51026759,"test":[0.6723155898]},
-{"learn":[0.6686399521],"iteration":163,"passed_time":6.108511297,"remaining_time":68.38552891,"test":[0.6722970834]},
-{"learn":[0.6686058279],"iteration":164,"passed_time":6.140719309,"remaining_time":68.29224202,"test":[0.6722872244]},
-{"learn":[0.6685761282],"iteration":165,"passed_time":6.169540017,"remaining_time":68.16226742,"test":[0.6722800481]},
-{"learn":[0.6685469327],"iteration":166,"passed_time":6.2020892,"remaining_time":68.07442817,"test":[0.6722550973]},
-{"learn":[0.6685157003],"iteration":167,"passed_time":6.231576547,"remaining_time":67.95385854,"test":[0.6722394313]},
-{"learn":[0.6684805143],"iteration":168,"passed_time":6.263261652,"remaining_time":67.85817802,"test":[0.6722204135]},
-{"learn":[0.6684485765],"iteration":169,"passed_time":6.295102833,"remaining_time":67.7649305,"test":[0.6721982148]},
-{"learn":[0.6684144429],"iteration":170,"passed_time":6.325415964,"remaining_time":67.65605729,"test":[0.6721971176]},
-{"learn":[0.6683849752],"iteration":171,"passed_time":6.35697084,"remaining_time":67.56129474,"test":[0.6721880705]},
-{"learn":[0.6683568537],"iteration":172,"passed_time":6.395913563,"remaining_time":67.5452837,"test":[0.672179176]},
-{"learn":[0.6683266628],"iteration":173,"passed_time":6.437330522,"remaining_time":67.55497433,"test":[0.6721769709]},
-{"learn":[0.6682937842],"iteration":174,"passed_time":6.472195712,"remaining_time":67.49575528,"test":[0.6721693215]},
-{"learn":[0.6682657097],"iteration":175,"passed_time":6.503044842,"remaining_time":67.395192,"test":[0.6721581386]},
-{"learn":[0.6682301443],"iteration":176,"passed_time":6.533528251,"remaining_time":67.29164972,"test":[0.6721638661]},
-{"learn":[0.6681995916],"iteration":177,"passed_time":6.562589882,"remaining_time":67.17437509,"test":[0.6721598475]},
-{"learn":[0.6681658267],"iteration":178,"passed_time":6.590816982,"remaining_time":67.04959623,"test":[0.6721433342]},
-{"learn":[0.6681422687],"iteration":179,"passed_time":6.624646227,"remaining_time":66.98253407,"test":[0.6721335599]},
-{"learn":[0.6681216601],"iteration":180,"passed_time":6.655147334,"remaining_time":66.88239227,"test":[0.6721300594]},
-{"learn":[0.6680899019],"iteration":181,"passed_time":6.687788902,"remaining_time":66.80439684,"test":[0.6721153533]},
-{"learn":[0.6680676394],"iteration":182,"passed_time":6.718057043,"remaining_time":66.7033314,"test":[0.6721076397]},
-{"learn":[0.6680413672],"iteration":183,"passed_time":6.751300957,"remaining_time":66.6324051,"test":[0.6721009911]},
-{"learn":[0.6680088406],"iteration":184,"passed_time":6.784288393,"remaining_time":66.55936991,"test":[0.6720999252]},
-{"learn":[0.6679873982],"iteration":185,"passed_time":6.810905309,"remaining_time":66.42463565,"test":[0.6720953028]},
-{"learn":[0.6679663544],"iteration":186,"passed_time":6.832974292,"remaining_time":66.24696466,"test":[0.6720942505]},
-{"learn":[0.6679417375],"iteration":187,"passed_time":6.867184511,"remaining_time":66.18796986,"test":[0.6720856237]},
-{"learn":[0.6679100197],"iteration":188,"passed_time":6.918652024,"remaining_time":66.29459691,"test":[0.6720876136]},
-{"learn":[0.667881208],"iteration":189,"passed_time":6.96948149,"remaining_time":66.39348156,"test":[0.6720880182]},
-{"learn":[0.6678475427],"iteration":190,"passed_time":7.018176318,"remaining_time":66.47058094,"test":[0.6720743856]},
-{"learn":[0.6678310341],"iteration":191,"passed_time":7.074099623,"remaining_time":66.61443812,"test":[0.6720598415]},
-{"learn":[0.6678060257],"iteration":192,"passed_time":7.117099742,"remaining_time":66.63522919,"test":[0.6720563492]},
-{"learn":[0.6677789336],"iteration":193,"passed_time":7.191058554,"remaining_time":66.94356571,"test":[0.6720389527]},
-{"learn":[0.6677478773],"iteration":194,"passed_time":7.2421897,"remaining_time":67.03667902,"test":[0.6720317324]},
-{"learn":[0.6677212408],"iteration":195,"passed_time":7.282401129,"remaining_time":67.02781447,"test":[0.672000736]},
-{"learn":[0.667704316],"iteration":196,"passed_time":7.317019235,"remaining_time":66.96744,"test":[0.6719895017]},
-{"learn":[0.6676819639],"iteration":197,"passed_time":7.351194179,"remaining_time":66.90329248,"test":[0.6719725302]},
-{"learn":[0.6676554448],"iteration":198,"passed_time":7.389840926,"remaining_time":66.87991712,"test":[0.6719770493]},
-{"learn":[0.6676318346],"iteration":199,"passed_time":7.432994652,"remaining_time":66.89695187,"test":[0.6719667172]},
-{"learn":[0.6676074705],"iteration":200,"passed_time":7.471295231,"remaining_time":66.86995085,"test":[0.6719511616]},
-{"learn":[0.6675849784],"iteration":201,"passed_time":7.506377837,"remaining_time":66.8141948,"test":[0.6719427289]},
-{"learn":[0.6675631744],"iteration":202,"passed_time":7.540821494,"remaining_time":66.75298633,"test":[0.6719299116]},
-{"learn":[0.6675397619],"iteration":203,"passed_time":7.56808212,"remaining_time":66.62880141,"test":[0.6719106583]},
-{"learn":[0.6675169086],"iteration":204,"passed_time":7.605676901,"remaining_time":66.59604896,"test":[0.6718967065]},
-{"learn":[0.6674864762],"iteration":205,"passed_time":7.638300222,"remaining_time":66.51995436,"test":[0.671890967]},
-{"learn":[0.6674670714],"iteration":206,"passed_time":7.665554951,"remaining_time":66.39777791,"test":[0.6718896293]},
-{"learn":[0.6674375599],"iteration":207,"passed_time":7.700277678,"remaining_time":66.34085384,"test":[0.6718883534]},
-{"learn":[0.6674148457],"iteration":208,"passed_time":7.734145802,"remaining_time":66.27681881,"test":[0.6718827289]},
-{"learn":[0.6673974446],"iteration":209,"passed_time":7.766232144,"remaining_time":66.19788351,"test":[0.6718763224]},
-{"learn":[0.6673812139],"iteration":210,"passed_time":7.796801222,"remaining_time":66.1065279,"test":[0.67187262]},
-{"learn":[0.6673515687],"iteration":211,"passed_time":7.831891449,"remaining_time":66.05387693,"test":[0.6718590402]},
-{"learn":[0.6673197956],"iteration":212,"passed_time":7.871259964,"remaining_time":66.0372843,"test":[0.6718455115]},
-{"learn":[0.6672900754],"iteration":213,"passed_time":7.910110502,"remaining_time":66.01615587,"test":[0.6718253747]},
-{"learn":[0.6672550009],"iteration":214,"passed_time":7.951342226,"remaining_time":66.01463197,"test":[0.671794877]},
-{"learn":[0.6672271563],"iteration":215,"passed_time":7.989001461,"remaining_time":65.98323429,"test":[0.6717873786]},
-{"learn":[0.667204521],"iteration":216,"passed_time":8.025973631,"remaining_time":65.94613357,"test":[0.6717765089]},
-{"learn":[0.667181968],"iteration":217,"passed_time":8.058434478,"remaining_time":65.87215707,"test":[0.6717616726]},
-{"learn":[0.6671640023],"iteration":218,"passed_time":8.087145957,"remaining_time":65.76806826,"test":[0.6717499215]},
-{"learn":[0.66714351],"iteration":219,"passed_time":8.112590578,"remaining_time":65.63823286,"test":[0.6717326052]},
-{"learn":[0.6671167156],"iteration":220,"passed_time":8.148644349,"remaining_time":65.59474342,"test":[0.6717161937]},
-{"learn":[0.6670915937],"iteration":221,"passed_time":8.197662625,"remaining_time":65.65515382,"test":[0.6717056951]},
-{"learn":[0.6670595279],"iteration":222,"passed_time":8.239228431,"remaining_time":65.65519696,"test":[0.6717021438]},
-{"learn":[0.667033994],"iteration":223,"passed_time":8.268371203,"remaining_time":65.55637168,"test":[0.6716868488]},
-{"learn":[0.6670008246],"iteration":224,"passed_time":8.298555216,"remaining_time":65.46638004,"test":[0.6716751909]},
-{"learn":[0.6669858319],"iteration":225,"passed_time":8.327401394,"remaining_time":65.36641625,"test":[0.671670116]},
-{"learn":[0.6669553964],"iteration":226,"passed_time":8.357648377,"remaining_time":65.27802014,"test":[0.6716558757]},
-{"learn":[0.6669274683],"iteration":227,"passed_time":8.384989701,"remaining_time":65.16755154,"test":[0.6716559962]},
-{"learn":[0.666896348],"iteration":228,"passed_time":8.418297538,"remaining_time":65.1039517,"test":[0.6716487875]},
-{"learn":[0.6668698686],"iteration":229,"passed_time":8.453919972,"remaining_time":65.05842761,"test":[0.6716427451]},
-{"learn":[0.6668513411],"iteration":230,"passed_time":8.49049033,"remaining_time":65.02024846,"test":[0.6716323255]},
-{"learn":[0.6668309985],"iteration":231,"passed_time":8.523986676,"remaining_time":64.95865708,"test":[0.6716303547]},
-{"learn":[0.6668058585],"iteration":232,"passed_time":8.550998228,"remaining_time":64.84812819,"test":[0.6716309509]},
-{"learn":[0.6667845908],"iteration":233,"passed_time":8.575382398,"remaining_time":64.71848425,"test":[0.6716215401]},
-{"learn":[0.6667582863],"iteration":234,"passed_time":8.607602961,"remaining_time":64.64859245,"test":[0.6716162103]},
-{"learn":[0.6667332943],"iteration":235,"passed_time":8.6353786,"remaining_time":64.54579597,"test":[0.6716135097]},
-{"learn":[0.6667070085],"iteration":236,"passed_time":8.66085309,"remaining_time":64.42651476,"test":[0.6716156696]},
-{"learn":[0.6666907315],"iteration":237,"passed_time":8.691362456,"remaining_time":64.34529684,"test":[0.6716020054]},
-{"learn":[0.6666633028],"iteration":238,"passed_time":8.719983169,"remaining_time":64.25058728,"test":[0.6715921704]},
-{"learn":[0.6666406707],"iteration":239,"passed_time":8.746012652,"remaining_time":64.13742611,"test":[0.6715804466]},
-{"learn":[0.6666134624],"iteration":240,"passed_time":8.773898765,"remaining_time":64.03853912,"test":[0.6715882966]},
-{"learn":[0.6665850522],"iteration":241,"passed_time":8.803292064,"remaining_time":63.9511878,"test":[0.6715753942]},
-{"learn":[0.6665631193],"iteration":242,"passed_time":8.833976809,"remaining_time":63.87365125,"test":[0.6715752261]},
-{"learn":[0.6665412643],"iteration":243,"passed_time":8.862338006,"remaining_time":63.7797768,"test":[0.6715625509]},
-{"learn":[0.6665168385],"iteration":244,"passed_time":8.892424073,"remaining_time":63.69879285,"test":[0.6715628214]},
-{"learn":[0.6664904845],"iteration":245,"passed_time":8.932383667,"remaining_time":63.68862175,"test":[0.6715601629]},
-{"learn":[0.6664678274],"iteration":246,"passed_time":8.962911123,"remaining_time":63.61126801,"test":[0.6715576255]},
-{"learn":[0.6664539777],"iteration":247,"passed_time":8.991624872,"remaining_time":63.52147894,"test":[0.6715550274]},
-{"learn":[0.6664334121],"iteration":248,"passed_time":9.021847081,"remaining_time":63.44278811,"test":[0.6715448645]},
-{"learn":[0.6664121724],"iteration":249,"passed_time":9.05121341,"remaining_time":63.35849387,"test":[0.6715308166]},
-{"learn":[0.666392034],"iteration":250,"passed_time":9.085113431,"remaining_time":63.30622865,"test":[0.671519334]},
-{"learn":[0.666366899],"iteration":251,"passed_time":9.110250512,"remaining_time":63.19332498,"test":[0.6715184071]},
-{"learn":[0.6663414098],"iteration":252,"passed_time":9.137253573,"remaining_time":63.09399997,"test":[0.6715163019]},
-{"learn":[0.6663157816],"iteration":253,"passed_time":9.174559864,"remaining_time":63.06606899,"test":[0.6715096094]},
-{"learn":[0.6662989799],"iteration":254,"passed_time":9.196898204,"remaining_time":62.93563673,"test":[0.6714992963]},
-{"learn":[0.6662696102],"iteration":255,"passed_time":9.238149902,"remaining_time":62.9348962,"test":[0.6714917256]},
-{"learn":[0.6662479711],"iteration":256,"passed_time":9.267818291,"remaining_time":62.85528125,"test":[0.671477406]},
-{"learn":[0.6662231874],"iteration":257,"passed_time":9.297538986,"remaining_time":62.77640665,"test":[0.6714741542]},
-{"learn":[0.6661947927],"iteration":258,"passed_time":9.324772701,"remaining_time":62.68119411,"test":[0.6714576155]},
-{"learn":[0.6661669951],"iteration":259,"passed_time":9.357824574,"remaining_time":62.62544138,"test":[0.6714473645]},
-{"learn":[0.6661426137],"iteration":260,"passed_time":9.388345461,"remaining_time":62.55299907,"test":[0.6714427232]},
-{"learn":[0.6661216749],"iteration":261,"passed_time":9.427290804,"remaining_time":62.53676114,"test":[0.6714364275]},
-{"learn":[0.6660983123],"iteration":262,"passed_time":9.461913185,"remaining_time":62.49179925,"test":[0.6714339587]},
-{"learn":[0.6660803402],"iteration":263,"passed_time":9.496090562,"remaining_time":62.44398945,"test":[0.6714336287]},
-{"learn":[0.6660617842],"iteration":264,"passed_time":9.524189317,"remaining_time":62.35648477,"test":[0.6714283568]},
-{"learn":[0.6660443878],"iteration":265,"passed_time":9.55372419,"remaining_time":62.27878852,"test":[0.6714271895]},
-{"learn":[0.6660176079],"iteration":266,"passed_time":9.590356068,"remaining_time":62.2475171,"test":[0.671413471]},
-{"learn":[0.6659967546],"iteration":267,"passed_time":9.620235131,"remaining_time":62.17256436,"test":[0.6714072396]},
-{"learn":[0.6659751467],"iteration":268,"passed_time":9.645948482,"remaining_time":62.0711406,"test":[0.6714002677]},
-{"learn":[0.6659539329],"iteration":269,"passed_time":9.682675077,"remaining_time":62.04084401,"test":[0.6714001163]},
-{"learn":[0.6659263951],"iteration":270,"passed_time":9.711914203,"remaining_time":61.96272936,"test":[0.6713933952]},
-{"learn":[0.6659038921],"iteration":271,"passed_time":9.739142426,"remaining_time":61.87219894,"test":[0.6713926761]},
-{"learn":[0.6658767418],"iteration":272,"passed_time":9.768751964,"remaining_time":61.79719649,"test":[0.6713836619]},
-{"learn":[0.6658510507],"iteration":273,"passed_time":9.804576737,"remaining_time":61.76167682,"test":[0.6713772112]},
-{"learn":[0.6658210119],"iteration":274,"passed_time":9.848653906,"remaining_time":61.77791996,"test":[0.6713603715]},
-{"learn":[0.6657963011],"iteration":275,"passed_time":9.88663261,"remaining_time":61.75563268,"test":[0.6713560246]},
-{"learn":[0.6657748552],"iteration":276,"passed_time":9.925808942,"remaining_time":61.74068161,"test":[0.6713837913]},
-{"learn":[0.6657490013],"iteration":277,"passed_time":9.965409489,"remaining_time":61.72818396,"test":[0.6713684274]},
-{"learn":[0.665732402],"iteration":278,"passed_time":9.99537326,"remaining_time":61.65604796,"test":[0.6713619356]},
-{"learn":[0.6657118786],"iteration":279,"passed_time":10.02216777,"remaining_time":61.5647449,"test":[0.6713584836]},
-{"learn":[0.665684467],"iteration":280,"passed_time":10.05593393,"remaining_time":61.51654955,"test":[0.6713673572]},
-{"learn":[0.6656584634],"iteration":281,"passed_time":10.08025153,"remaining_time":61.41089406,"test":[0.6713625568]},
-{"learn":[0.6656309991],"iteration":282,"passed_time":10.11102202,"remaining_time":61.34496401,"test":[0.6713542652]},
-{"learn":[0.6656073482],"iteration":283,"passed_time":10.14714598,"remaining_time":61.31162855,"test":[0.6713512017]},
-{"learn":[0.6655890957],"iteration":284,"passed_time":10.17528061,"remaining_time":61.23019734,"test":[0.671342038]},
-{"learn":[0.6655665563],"iteration":285,"passed_time":10.2021403,"remaining_time":61.14149818,"test":[0.6713279798]},
-{"learn":[0.6655452454],"iteration":286,"passed_time":10.23423432,"remaining_time":61.08447174,"test":[0.6713123285]},
-{"learn":[0.6655255286],"iteration":287,"passed_time":10.26481698,"remaining_time":61.0186343,"test":[0.6713035326]},
-{"learn":[0.6655053548],"iteration":288,"passed_time":10.29945844,"remaining_time":60.97707056,"test":[0.6713022203]},
-{"learn":[0.6654893396],"iteration":289,"passed_time":10.32366496,"remaining_time":60.87402441,"test":[0.671296041]},
-{"learn":[0.6654648912],"iteration":290,"passed_time":10.35344703,"remaining_time":60.80426453,"test":[0.6712829551]},
-{"learn":[0.6654442759],"iteration":291,"passed_time":10.3949915,"remaining_time":60.8035804,"test":[0.6712769751]},
-{"learn":[0.6654173127],"iteration":292,"passed_time":10.43148765,"remaining_time":60.77320621,"test":[0.6712702915]},
-{"learn":[0.6653914518],"iteration":293,"passed_time":10.47162738,"remaining_time":60.76393303,"test":[0.6712379343]},
-{"learn":[0.6653648946],"iteration":294,"passed_time":10.50360107,"remaining_time":60.70725362,"test":[0.6712192006]},
-{"learn":[0.665344141],"iteration":295,"passed_time":10.53460819,"remaining_time":60.64517686,"test":[0.6712074061]},
-{"learn":[0.6653140817],"iteration":296,"passed_time":10.57659448,"remaining_time":60.64626395,"test":[0.6711953324]},
-{"learn":[0.665295365],"iteration":297,"passed_time":10.61260262,"remaining_time":60.61291829,"test":[0.6711891001]},
-{"learn":[0.6652787488],"iteration":298,"passed_time":10.63910358,"remaining_time":60.52546889,"test":[0.6711870526]},
-{"learn":[0.6652502991],"iteration":299,"passed_time":10.6681867,"remaining_time":60.45305797,"test":[0.6711812809]},
-{"learn":[0.665231168],"iteration":300,"passed_time":10.70260503,"remaining_time":60.41104967,"test":[0.6711768946]},
-{"learn":[0.6652136682],"iteration":301,"passed_time":10.72952096,"remaining_time":60.32690925,"test":[0.6711845012]},
-{"learn":[0.6651903001],"iteration":302,"passed_time":10.76489952,"remaining_time":60.29054288,"test":[0.6711869636]},
-{"learn":[0.6651697153],"iteration":303,"passed_time":10.80197155,"remaining_time":60.26363073,"test":[0.671186884]},
-{"learn":[0.6651525958],"iteration":304,"passed_time":10.82922271,"remaining_time":60.18207375,"test":[0.6711890401]},
-{"learn":[0.6651322685],"iteration":305,"passed_time":10.8578399,"remaining_time":60.10843394,"test":[0.6711868603]},
-{"learn":[0.6651113828],"iteration":306,"passed_time":10.89228879,"remaining_time":60.06724727,"test":[0.6711900892]},
-{"learn":[0.6650886807],"iteration":307,"passed_time":10.93056436,"remaining_time":60.04712628,"test":[0.6711884242]},
-{"learn":[0.6650622251],"iteration":308,"passed_time":10.97231236,"remaining_time":60.04589061,"test":[0.6711837119]},
-{"learn":[0.6650429987],"iteration":309,"passed_time":11.00296848,"remaining_time":59.98392494,"test":[0.6711766645]},
-{"learn":[0.665015513],"iteration":310,"passed_time":11.03002276,"remaining_time":59.90259947,"test":[0.671172959]},
-{"learn":[0.6650019022],"iteration":311,"passed_time":11.05828865,"remaining_time":59.82817707,"test":[0.6711740433]},
-{"learn":[0.664979951],"iteration":312,"passed_time":11.09287745,"remaining_time":59.78812863,"test":[0.6711715069]},
-{"learn":[0.6649549638],"iteration":313,"passed_time":11.1177757,"remaining_time":59.69608229,"test":[0.6711589843]},
-{"learn":[0.6649340455],"iteration":314,"passed_time":11.14959087,"remaining_time":59.64146228,"test":[0.6711446402]},
-{"learn":[0.6649162445],"iteration":315,"passed_time":11.18718772,"remaining_time":59.61779784,"test":[0.6711415366]},
-{"learn":[0.6649048119],"iteration":316,"passed_time":11.21179073,"remaining_time":59.52505932,"test":[0.6711359351]},
-{"learn":[0.6648796463],"iteration":317,"passed_time":11.24311165,"remaining_time":59.46828238,"test":[0.671143361]},
-{"learn":[0.6648605481],"iteration":318,"passed_time":11.27486028,"remaining_time":59.41391889,"test":[0.6711353638]},
-{"learn":[0.6648429084],"iteration":319,"passed_time":11.30400807,"remaining_time":59.34604237,"test":[0.6711444387]},
-{"learn":[0.6648238121],"iteration":320,"passed_time":11.33488419,"remaining_time":59.28744721,"test":[0.6711487352]},
-{"learn":[0.6647969527],"iteration":321,"passed_time":11.36208838,"remaining_time":59.20988915,"test":[0.67114436]},
-{"learn":[0.6647854723],"iteration":322,"passed_time":11.39429642,"remaining_time":59.15862259,"test":[0.6711444722]},
-{"learn":[0.6647589304],"iteration":323,"passed_time":11.4363998,"remaining_time":59.15866068,"test":[0.6711325635]},
-{"learn":[0.6647429024],"iteration":324,"passed_time":11.47751019,"remaining_time":59.15332173,"test":[0.6711269403]},
-{"learn":[0.6647237508],"iteration":325,"passed_time":11.5136833,"remaining_time":59.12241054,"test":[0.6711154078]},
-{"learn":[0.6647059396],"iteration":326,"passed_time":11.54795566,"remaining_time":59.08174257,"test":[0.6711203043]},
-{"learn":[0.664686288],"iteration":327,"passed_time":11.57245915,"remaining_time":58.99131613,"test":[0.6711241333]},
-{"learn":[0.6646532527],"iteration":328,"passed_time":11.60790333,"remaining_time":58.95685857,"test":[0.6711213497]},
-{"learn":[0.6646306438],"iteration":329,"passed_time":11.63787346,"remaining_time":58.89469298,"test":[0.6711231641]},
-{"learn":[0.6646098516],"iteration":330,"passed_time":11.66805718,"remaining_time":58.83379887,"test":[0.6711049215]},
-{"learn":[0.6645858284],"iteration":331,"passed_time":11.70070223,"remaining_time":58.78545579,"test":[0.6711031963]},
-{"learn":[0.6645707188],"iteration":332,"passed_time":11.724753,"remaining_time":58.69418391,"test":[0.6710996314]},
-{"learn":[0.6645485788],"iteration":333,"passed_time":11.75795297,"remaining_time":58.64895104,"test":[0.6710867309]},
-{"learn":[0.6645305696],"iteration":334,"passed_time":11.78053066,"remaining_time":58.55099567,"test":[0.6710914578]},
-{"learn":[0.6645108881],"iteration":335,"passed_time":11.81570271,"remaining_time":58.51586106,"test":[0.6710929585]},
-{"learn":[0.6644923286],"iteration":336,"passed_time":11.8448851,"remaining_time":58.45116888,"test":[0.6710984779]},
-{"learn":[0.6644805222],"iteration":337,"passed_time":11.86964023,"remaining_time":58.36491734,"test":[0.6710923199]},
-{"learn":[0.6644572776],"iteration":338,"passed_time":11.90591446,"remaining_time":58.33546879,"test":[0.6710893917]},
-{"learn":[0.6644320741],"iteration":339,"passed_time":11.94145444,"remaining_time":58.30239521,"test":[0.6710923306]},
-{"learn":[0.6644115048],"iteration":340,"passed_time":11.98658051,"remaining_time":58.31594449,"test":[0.6710927901]},
-{"learn":[0.6643949013],"iteration":341,"passed_time":12.02038848,"remaining_time":58.27428098,"test":[0.6711092802]},
-{"learn":[0.6643619789],"iteration":342,"passed_time":12.06653941,"remaining_time":58.29229096,"test":[0.6711012995]},
-{"learn":[0.6643389502],"iteration":343,"passed_time":12.12283646,"remaining_time":58.35877087,"test":[0.6711015305]},
-{"learn":[0.6643088915],"iteration":344,"passed_time":12.17733618,"remaining_time":58.41591705,"test":[0.6710975574]},
-{"learn":[0.664286972],"iteration":345,"passed_time":12.22133732,"remaining_time":58.42223099,"test":[0.6710899474]},
-{"learn":[0.664274149],"iteration":346,"passed_time":12.2642467,"remaining_time":58.42305415,"test":[0.671085152]},
-{"learn":[0.6642536926],"iteration":347,"passed_time":12.30091895,"remaining_time":58.39401755,"test":[0.6710814533]},
-{"learn":[0.6642357634],"iteration":348,"passed_time":12.32484094,"remaining_time":58.30462002,"test":[0.6710701892]},
-{"learn":[0.664207914],"iteration":349,"passed_time":12.35469303,"remaining_time":58.24355287,"test":[0.67105503]},
-{"learn":[0.6641853097],"iteration":350,"passed_time":12.40148755,"remaining_time":58.26225919,"test":[0.6710527861]},
-{"learn":[0.6641654917],"iteration":351,"passed_time":12.43803877,"remaining_time":58.23263605,"test":[0.6710508715]},
-{"learn":[0.664143804],"iteration":352,"passed_time":12.47995438,"remaining_time":58.22800245,"test":[0.6710560803]},
-{"learn":[0.6641290647],"iteration":353,"passed_time":12.51241326,"remaining_time":58.17918707,"test":[0.6710465693]},
-{"learn":[0.6641117244],"iteration":354,"passed_time":12.5417829,"remaining_time":58.11614893,"test":[0.6710440741]},
-{"learn":[0.6640880219],"iteration":355,"passed_time":12.5692936,"remaining_time":58.0447154,"test":[0.6710496913]},
-{"learn":[0.6640669415],"iteration":356,"passed_time":12.5976392,"remaining_time":57.97737034,"test":[0.6710404659]},
-{"learn":[0.6640462999],"iteration":357,"passed_time":12.62815847,"remaining_time":57.92021287,"test":[0.6710293986]},
-{"learn":[0.664030296],"iteration":358,"passed_time":12.65342509,"remaining_time":57.8391938,"test":[0.6710353817]},
-{"learn":[0.6640028542],"iteration":359,"passed_time":12.68233453,"remaining_time":57.77507954,"test":[0.6710271815]},
-{"learn":[0.6639813347],"iteration":360,"passed_time":12.72037964,"remaining_time":57.75263774,"test":[0.6710288077]},
-{"learn":[0.6639597941],"iteration":361,"passed_time":12.744473,"remaining_time":57.66698004,"test":[0.6710169894]},
-{"learn":[0.6639429832],"iteration":362,"passed_time":12.77086568,"remaining_time":57.59203063,"test":[0.6710119848]},
-{"learn":[0.6639222708],"iteration":363,"passed_time":12.81194554,"remaining_time":57.58335961,"test":[0.6710114775]},
-{"learn":[0.6639065546],"iteration":364,"passed_time":12.84133287,"remaining_time":57.52213492,"test":[0.6710013614]},
-{"learn":[0.6638823236],"iteration":365,"passed_time":12.87057337,"remaining_time":57.46042866,"test":[0.6709985657]},
-{"learn":[0.6638648195],"iteration":366,"passed_time":12.8971183,"remaining_time":57.38690512,"test":[0.6709948954]},
-{"learn":[0.6638436235],"iteration":367,"passed_time":12.93825161,"remaining_time":57.37833324,"test":[0.6709970591]},
-{"learn":[0.6638208732],"iteration":368,"passed_time":12.97444296,"remaining_time":57.3477411,"test":[0.6709739289]},
-{"learn":[0.6637956357],"iteration":369,"passed_time":13.00974924,"remaining_time":57.31321963,"test":[0.6709754911]},
-{"learn":[0.6637718453],"iteration":370,"passed_time":13.03832239,"remaining_time":57.24912984,"test":[0.6709717066]},
-{"learn":[0.663756918],"iteration":371,"passed_time":13.07843077,"remaining_time":57.23571316,"test":[0.67096845]},
-{"learn":[0.6637353525],"iteration":372,"passed_time":13.11729124,"remaining_time":57.21671005,"test":[0.6709739445]},
-{"learn":[0.6637143112],"iteration":373,"passed_time":13.14745329,"remaining_time":57.15978354,"test":[0.6709728881]},
-{"learn":[0.6636956547],"iteration":374,"passed_time":13.18118022,"remaining_time":57.11844761,"test":[0.6709694284]},
-{"learn":[0.663680995],"iteration":375,"passed_time":13.20539229,"remaining_time":57.03605604,"test":[0.6709604166]},
-{"learn":[0.66366728],"iteration":376,"passed_time":13.23563977,"remaining_time":56.97995583,"test":[0.6709605025]},
-{"learn":[0.6636487567],"iteration":377,"passed_time":13.27428255,"remaining_time":56.96001665,"test":[0.6709603727]},
-{"learn":[0.6636266904],"iteration":378,"passed_time":13.30625754,"remaining_time":56.91146033,"test":[0.670944339]},
-{"learn":[0.6636116064],"iteration":379,"passed_time":13.33327871,"remaining_time":56.84187241,"test":[0.6709447187]},
-{"learn":[0.6635902746],"iteration":380,"passed_time":13.36632239,"remaining_time":56.79809961,"test":[0.6709538679]},
-{"learn":[0.6635654896],"iteration":381,"passed_time":13.39639051,"remaining_time":56.74177969,"test":[0.6709640912]},
-{"learn":[0.6635393029],"iteration":382,"passed_time":13.42189438,"remaining_time":56.66632694,"test":[0.6709534847]},
-{"learn":[0.6635171734],"iteration":383,"passed_time":13.46730432,"remaining_time":56.6749057,"test":[0.6709471555]},
-{"learn":[0.663500789],"iteration":384,"passed_time":13.50832777,"remaining_time":56.66480351,"test":[0.6709506783]},
-{"learn":[0.663477743],"iteration":385,"passed_time":13.54029627,"remaining_time":56.61667921,"test":[0.6709546729]},
-{"learn":[0.6634584806],"iteration":386,"passed_time":13.56996301,"remaining_time":56.5590448,"test":[0.670930774]},
-{"learn":[0.6634337499],"iteration":387,"passed_time":13.59835745,"remaining_time":56.4962686,"test":[0.6709287322]},
-{"learn":[0.6634135584],"iteration":388,"passed_time":13.6279617,"remaining_time":56.43867943,"test":[0.6709198643]},
-{"learn":[0.6633868455],"iteration":389,"passed_time":13.65633448,"remaining_time":56.37615005,"test":[0.6709220389]},
-{"learn":[0.6633755323],"iteration":390,"passed_time":13.68565529,"remaining_time":56.31769658,"test":[0.6709230923]},
-{"learn":[0.663356103],"iteration":391,"passed_time":13.71789303,"remaining_time":56.27135714,"test":[0.670930414]},
-{"learn":[0.6633337631],"iteration":392,"passed_time":13.75060752,"remaining_time":56.2270389,"test":[0.6709354296]},
-{"learn":[0.663319422],"iteration":393,"passed_time":13.77167974,"remaining_time":56.13532403,"test":[0.6709351544]},
-{"learn":[0.6632911566],"iteration":394,"passed_time":13.80416242,"remaining_time":56.09033084,"test":[0.6709414935]},
-{"learn":[0.6632687875],"iteration":395,"passed_time":13.82525369,"remaining_time":55.9992599,"test":[0.6709445943]},
-{"learn":[0.6632431997],"iteration":396,"passed_time":13.85836516,"remaining_time":55.95707646,"test":[0.6709475685]},
-{"learn":[0.6632189331],"iteration":397,"passed_time":13.88898168,"remaining_time":55.90489613,"test":[0.6709533591]},
-{"learn":[0.663201035],"iteration":398,"passed_time":13.91726355,"remaining_time":55.84345598,"test":[0.6709592222]},
-{"learn":[0.6631898553],"iteration":399,"passed_time":13.95316828,"remaining_time":55.81267311,"test":[0.6709508704]},
-{"learn":[0.6631712482],"iteration":400,"passed_time":13.99418497,"remaining_time":55.80224881,"test":[0.6709479912]},
-{"learn":[0.663143025],"iteration":401,"passed_time":14.0253575,"remaining_time":55.75254052,"test":[0.6709417519]},
-{"learn":[0.663121538],"iteration":402,"passed_time":14.04844239,"remaining_time":55.67087467,"test":[0.6709476082]},
-{"learn":[0.6631087792],"iteration":403,"passed_time":14.0761289,"remaining_time":55.60767753,"test":[0.6709480979]},
-{"learn":[0.6630859067],"iteration":404,"passed_time":14.10555105,"remaining_time":55.55149118,"test":[0.6709448724]},
-{"learn":[0.663066483],"iteration":405,"passed_time":14.1427661,"remaining_time":55.52603242,"test":[0.6709421934]},
-{"learn":[0.6630443652],"iteration":406,"passed_time":14.18285552,"remaining_time":55.51176619,"test":[0.6709386261]},
-{"learn":[0.6630250376],"iteration":407,"passed_time":14.21458769,"remaining_time":55.46476372,"test":[0.6709461564]},
-{"learn":[0.6630007822],"iteration":408,"passed_time":14.24035708,"remaining_time":55.39464088,"test":[0.670934384]},
-{"learn":[0.6629768728],"iteration":409,"passed_time":14.26711915,"remaining_time":55.32858403,"test":[0.6709312987]},
-{"learn":[0.6629528093],"iteration":410,"passed_time":14.29943785,"remaining_time":55.28420133,"test":[0.670931806]},
-{"learn":[0.6629260936],"iteration":411,"passed_time":14.32489173,"remaining_time":55.21341763,"test":[0.6709286111]},
-{"learn":[0.6629102182],"iteration":412,"passed_time":14.35119075,"remaining_time":55.14610101,"test":[0.6709224729]},
-{"learn":[0.6628863488],"iteration":413,"passed_time":14.37946054,"remaining_time":55.08653242,"test":[0.6709236504]},
-{"learn":[0.6628648972],"iteration":414,"passed_time":14.41005914,"remaining_time":55.03600899,"test":[0.6709245901]},
-{"learn":[0.6628454339],"iteration":415,"passed_time":14.45103793,"remaining_time":55.02510598,"test":[0.6709463437]},
-{"learn":[0.6628200274],"iteration":416,"passed_time":14.48428995,"remaining_time":54.98472661,"test":[0.6709567049]},
-{"learn":[0.6627942591],"iteration":417,"passed_time":14.5135184,"remaining_time":54.92915339,"test":[0.670945606]},
-{"learn":[0.6627744647],"iteration":418,"passed_time":14.53698524,"remaining_time":54.85196578,"test":[0.6709479298]},
-{"learn":[0.662765485],"iteration":419,"passed_time":14.56542473,"remaining_time":54.79374067,"test":[0.6709464351]},
-{"learn":[0.6627503257],"iteration":420,"passed_time":14.58728594,"remaining_time":54.71098455,"test":[0.6709414048]},
-{"learn":[0.6627323029],"iteration":421,"passed_time":14.61501375,"remaining_time":54.65045425,"test":[0.6709414427]},
-{"learn":[0.6627111509],"iteration":422,"passed_time":14.64231614,"remaining_time":54.58849302,"test":[0.6709296343]},
-{"learn":[0.6626785863],"iteration":423,"passed_time":14.66665432,"remaining_time":54.51567739,"test":[0.670924721]},
-{"learn":[0.6626576561],"iteration":424,"passed_time":14.69050441,"remaining_time":54.44128104,"test":[0.670906284]},
-{"learn":[0.6626363113],"iteration":425,"passed_time":14.71910475,"remaining_time":54.38467341,"test":[0.6708996826]},
-{"learn":[0.6626181065],"iteration":426,"passed_time":14.73941058,"remaining_time":54.2976413,"test":[0.6708987677]},
-{"learn":[0.66259794],"iteration":427,"passed_time":14.77242451,"remaining_time":54.25759657,"test":[0.670909526]},
-{"learn":[0.6625765658],"iteration":428,"passed_time":14.79088688,"remaining_time":54.1642967,"test":[0.6709033226]},
-{"learn":[0.6625526572],"iteration":429,"passed_time":14.82430966,"remaining_time":54.12596783,"test":[0.6708750209]},
-{"learn":[0.66253135],"iteration":430,"passed_time":14.84439175,"remaining_time":54.03909666,"test":[0.6708752079]},
-{"learn":[0.6625035695],"iteration":431,"passed_time":14.8764415,"remaining_time":53.99597284,"test":[0.6708776566]},
-{"learn":[0.662480212],"iteration":432,"passed_time":14.90666075,"remaining_time":53.94627573,"test":[0.6708736133]},
-{"learn":[0.6624611632],"iteration":433,"passed_time":14.93845927,"remaining_time":53.90236684,"test":[0.6708754298]},
-{"learn":[0.6624332625],"iteration":434,"passed_time":14.98024104,"remaining_time":53.89443041,"test":[0.6708751084]},
-{"learn":[0.6624120584],"iteration":435,"passed_time":15.00605075,"remaining_time":53.82904442,"test":[0.6708642042]},
-{"learn":[0.6623941719],"iteration":436,"passed_time":15.03384083,"remaining_time":53.77092268,"test":[0.6708610465]},
-{"learn":[0.6623766304],"iteration":437,"passed_time":15.05972545,"remaining_time":53.70614417,"test":[0.6708574768]},
-{"learn":[0.6623623329],"iteration":438,"passed_time":15.08505889,"remaining_time":53.63958297,"test":[0.6708557953]},
-{"learn":[0.6623442925],"iteration":439,"passed_time":15.11080547,"remaining_time":53.57467393,"test":[0.670871378]},
-{"learn":[0.6623212715],"iteration":440,"passed_time":15.13466304,"remaining_time":53.50326458,"test":[0.6708640187]},
-{"learn":[0.6623025941],"iteration":441,"passed_time":15.16037021,"remaining_time":53.43859001,"test":[0.6708700565]},
-{"learn":[0.6622749791],"iteration":442,"passed_time":15.18471062,"remaining_time":53.36928767,"test":[0.6708667534]},
-{"learn":[0.6622534499],"iteration":443,"passed_time":15.21140556,"remaining_time":53.30843931,"test":[0.6708675383]},
-{"learn":[0.6622305473],"iteration":444,"passed_time":15.23498219,"remaining_time":53.23684787,"test":[0.6708740175]},
-{"learn":[0.6622059333],"iteration":445,"passed_time":15.26647355,"remaining_time":53.19304911,"test":[0.6708774523]},
-{"learn":[0.6621871707],"iteration":446,"passed_time":15.28793136,"remaining_time":53.11444609,"test":[0.6708697231]},
-{"learn":[0.6621638454],"iteration":447,"passed_time":15.31613827,"remaining_time":53.05947899,"test":[0.6708614971]},
-{"learn":[0.6621511296],"iteration":448,"passed_time":15.33689091,"remaining_time":52.9788815,"test":[0.6708607946]},
-{"learn":[0.6621349978],"iteration":449,"passed_time":15.36674634,"remaining_time":52.92990406,"test":[0.6708740865]},
-{"learn":[0.6621120424],"iteration":450,"passed_time":15.393642,"remaining_time":52.87084582,"test":[0.6708729562]},
-{"learn":[0.6620958271],"iteration":451,"passed_time":15.42984657,"remaining_time":52.84381082,"test":[0.6708674017]},
-{"learn":[0.6620793528],"iteration":452,"passed_time":15.46956188,"remaining_time":52.82872456,"test":[0.6708693088]},
-{"learn":[0.6620572713],"iteration":453,"passed_time":15.49032259,"remaining_time":52.74898396,"test":[0.6708712037]},
-{"learn":[0.6620395025],"iteration":454,"passed_time":15.52379393,"remaining_time":52.71266289,"test":[0.6708703905]},
-{"learn":[0.6620188044],"iteration":455,"passed_time":15.55053135,"remaining_time":52.65355352,"test":[0.6708577595]},
-{"learn":[0.6620017347],"iteration":456,"passed_time":15.57735398,"remaining_time":52.59487352,"test":[0.6708493546]},
-{"learn":[0.6619811454],"iteration":457,"passed_time":15.60434803,"remaining_time":52.53690973,"test":[0.6708523777]},
-{"learn":[0.6619695569],"iteration":458,"passed_time":15.63056555,"remaining_time":52.47647387,"test":[0.6708454134]},
-{"learn":[0.661952377],"iteration":459,"passed_time":15.656355,"remaining_time":52.41475368,"test":[0.6708404483]},
-{"learn":[0.6619237442],"iteration":460,"passed_time":15.68232112,"remaining_time":52.35377918,"test":[0.6708274771]},
-{"learn":[0.6619089407],"iteration":461,"passed_time":15.71164945,"remaining_time":52.30414904,"test":[0.6708244992]},
-{"learn":[0.6618886168],"iteration":462,"passed_time":15.7361944,"remaining_time":52.23872743,"test":[0.6708344314]},
-{"learn":[0.6618831383],"iteration":463,"passed_time":15.76527735,"remaining_time":52.18850433,"test":[0.6708279081]},
-{"learn":[0.6618690774],"iteration":464,"passed_time":15.78652262,"remaining_time":52.11249942,"test":[0.6708258106]},
-{"learn":[0.661845878],"iteration":465,"passed_time":15.81756836,"remaining_time":52.06899113,"test":[0.6708049714]},
-{"learn":[0.6618290213],"iteration":466,"passed_time":15.83979966,"remaining_time":51.99660146,"test":[0.670810989]},
-{"learn":[0.6618050064],"iteration":467,"passed_time":15.87342473,"remaining_time":51.9617237,"test":[0.6708212237]},
-{"learn":[0.6617832833],"iteration":468,"passed_time":15.90381555,"remaining_time":51.9162934,"test":[0.6708221741]},
-{"learn":[0.6617652311],"iteration":469,"passed_time":15.93502938,"remaining_time":51.87360627,"test":[0.6708259658]},
-{"learn":[0.6617443144],"iteration":470,"passed_time":15.96919221,"remaining_time":51.84054117,"test":[0.6708159692]},
-{"learn":[0.6617202619],"iteration":471,"passed_time":15.99477329,"remaining_time":51.77968981,"test":[0.6708136212]},
-{"learn":[0.6617005831],"iteration":472,"passed_time":16.02279091,"remaining_time":51.72685354,"test":[0.6708224942]},
-{"learn":[0.6616824419],"iteration":473,"passed_time":16.04763422,"remaining_time":51.66390258,"test":[0.6708363084]},
-{"learn":[0.6616538226],"iteration":474,"passed_time":16.07374645,"remaining_time":51.60518598,"test":[0.670850875]},
-{"learn":[0.6616314155],"iteration":475,"passed_time":16.09993591,"remaining_time":51.54685363,"test":[0.6708527236]},
-{"learn":[0.6616127861],"iteration":476,"passed_time":16.12811357,"remaining_time":51.49500411,"test":[0.6708453401]},
-{"learn":[0.6616029072],"iteration":477,"passed_time":16.15264086,"remaining_time":51.43163051,"test":[0.6708413844]},
-{"learn":[0.6615843751],"iteration":478,"passed_time":16.17696751,"remaining_time":51.36778201,"test":[0.6708364569]},
-{"learn":[0.661563216],"iteration":479,"passed_time":16.20551145,"remaining_time":51.31745293,"test":[0.6708251774]},
-{"learn":[0.6615432257],"iteration":480,"passed_time":16.22860577,"remaining_time":51.2500045,"test":[0.6708154393]},
-{"learn":[0.6615263324],"iteration":481,"passed_time":16.25544093,"remaining_time":51.19452144,"test":[0.6708111613]},
-{"learn":[0.6615033259],"iteration":482,"passed_time":16.27729221,"remaining_time":51.12350369,"test":[0.6708102339]},
-{"learn":[0.661484293],"iteration":483,"passed_time":16.30502335,"remaining_time":51.07110619,"test":[0.6707929623]},
-{"learn":[0.6614678231],"iteration":484,"passed_time":16.32842702,"remaining_time":51.00529266,"test":[0.6707900226]},
-{"learn":[0.6614463024],"iteration":485,"passed_time":16.36272839,"remaining_time":50.97360242,"test":[0.6707832384]},
-{"learn":[0.6614155436],"iteration":486,"passed_time":16.39272506,"remaining_time":50.92852776,"test":[0.6707739118]},
-{"learn":[0.6613958945],"iteration":487,"passed_time":16.42636604,"remaining_time":50.89480625,"test":[0.6707737538]},
-{"learn":[0.661380611],"iteration":488,"passed_time":16.4597142,"remaining_time":50.86018027,"test":[0.6707730234]},
-{"learn":[0.6613677802],"iteration":489,"passed_time":16.48056007,"remaining_time":50.78703206,"test":[0.6707796291]},
-{"learn":[0.6613530086],"iteration":490,"passed_time":16.51091177,"remaining_time":50.74331132,"test":[0.670791408]},
-{"learn":[0.6613248211],"iteration":491,"passed_time":16.53097438,"remaining_time":50.66810846,"test":[0.6707944906]},
-{"learn":[0.6613059359],"iteration":492,"passed_time":16.56161402,"remaining_time":50.62546112,"test":[0.6707835635]},
-{"learn":[0.6612729965],"iteration":493,"passed_time":16.5854633,"remaining_time":50.56216139,"test":[0.6707908928]},
-{"learn":[0.6612624948],"iteration":494,"passed_time":16.61302735,"remaining_time":50.51031547,"test":[0.670796262]},
-{"learn":[0.6612401679],"iteration":495,"passed_time":16.63896978,"remaining_time":50.45365029,"test":[0.6707877825]},
-{"learn":[0.6612191637],"iteration":496,"passed_time":16.663707,"remaining_time":50.39346403,"test":[0.6707854132]},
-{"learn":[0.6611912219],"iteration":497,"passed_time":16.69040179,"remaining_time":50.33932428,"test":[0.6707756206]},
-{"learn":[0.6611773017],"iteration":498,"passed_time":16.71612789,"remaining_time":50.28238068,"test":[0.6707707899]},
-{"learn":[0.6611638216],"iteration":499,"passed_time":16.74072553,"remaining_time":50.2221766,"test":[0.6707704386]},
-{"learn":[0.6611450533],"iteration":500,"passed_time":16.77346538,"remaining_time":50.18647626,"test":[0.6707621465]},
-{"learn":[0.6611179111],"iteration":501,"passed_time":16.80230735,"remaining_time":50.13915621,"test":[0.6707661931]},
-{"learn":[0.6610959069],"iteration":502,"passed_time":16.83637769,"remaining_time":50.10747,"test":[0.6707651988]},
-{"learn":[0.6610728788],"iteration":503,"passed_time":16.87382128,"remaining_time":50.08578697,"test":[0.6707607827]},
-{"learn":[0.6610436668],"iteration":504,"passed_time":16.92151611,"remaining_time":50.09438927,"test":[0.670760242]},
-{"learn":[0.6610188976],"iteration":505,"passed_time":16.9898618,"remaining_time":50.16374216,"test":[0.6707506008]},
-{"learn":[0.6610030555],"iteration":506,"passed_time":17.03818668,"remaining_time":50.17359509,"test":[0.6707452886]},
-{"learn":[0.6609831174],"iteration":507,"passed_time":17.06933058,"remaining_time":50.13275833,"test":[0.6707355189]},
-{"learn":[0.6609586562],"iteration":508,"passed_time":17.1106164,"remaining_time":50.12166807,"test":[0.6707312551]},
-{"learn":[0.660935882],"iteration":509,"passed_time":17.14537899,"remaining_time":50.09140137,"test":[0.6707199485]},
-{"learn":[0.6609202024],"iteration":510,"passed_time":17.19066307,"remaining_time":50.09177556,"test":[0.6707131947]},
-{"learn":[0.6609011137],"iteration":511,"passed_time":17.21958034,"remaining_time":50.04440537,"test":[0.6707154112]},
-{"learn":[0.6608726737],"iteration":512,"passed_time":17.24756917,"remaining_time":49.99441591,"test":[0.6706982346]},
-{"learn":[0.6608608849],"iteration":513,"passed_time":17.27150822,"remaining_time":49.93280391,"test":[0.6706988941]},
-{"learn":[0.6608387256],"iteration":514,"passed_time":17.29800365,"remaining_time":49.87870957,"test":[0.6706989098]},
-{"learn":[0.6608136063],"iteration":515,"passed_time":17.34332283,"remaining_time":49.87885868,"test":[0.670693306]},
-{"learn":[0.6607946343],"iteration":516,"passed_time":17.37393636,"remaining_time":49.83664916,"test":[0.6706944515]},
-{"learn":[0.6607703935],"iteration":517,"passed_time":17.4173655,"remaining_time":49.83114994,"test":[0.6706899688]},
-{"learn":[0.6607509625],"iteration":518,"passed_time":17.46008645,"remaining_time":49.82348368,"test":[0.6706909374]},
-{"learn":[0.6607238109],"iteration":519,"passed_time":17.4906988,"remaining_time":49.78121967,"test":[0.6706855074]},
-{"learn":[0.6606999858],"iteration":520,"passed_time":17.5186435,"remaining_time":49.7314275,"test":[0.6706787779]},
-{"learn":[0.6606813873],"iteration":521,"passed_time":17.54613056,"remaining_time":49.6804233,"test":[0.6706737082]},
-{"learn":[0.6606610372],"iteration":522,"passed_time":17.57100039,"remaining_time":49.62211774,"test":[0.6706761225]},
-{"learn":[0.660638456],"iteration":523,"passed_time":17.60084283,"remaining_time":49.5779466,"test":[0.670685455]},
-{"learn":[0.6606156483],"iteration":524,"passed_time":17.62599925,"remaining_time":49.52066456,"test":[0.6706693855]},
-{"learn":[0.6605968623],"iteration":525,"passed_time":17.65519625,"remaining_time":49.47482751,"test":[0.6706647216]},
-{"learn":[0.6605735776],"iteration":526,"passed_time":17.67910836,"remaining_time":49.41428199,"test":[0.6706569188]},
-{"learn":[0.6605517294],"iteration":527,"passed_time":17.70744827,"remaining_time":49.36621942,"test":[0.6706549134]},
-{"learn":[0.6605309239],"iteration":528,"passed_time":17.72943083,"remaining_time":49.3005534,"test":[0.6706547978]},
-{"learn":[0.6605086434],"iteration":529,"passed_time":17.75830336,"remaining_time":49.25416215,"test":[0.6706564214]},
-{"learn":[0.6604803349],"iteration":530,"passed_time":17.78141858,"remaining_time":49.19190939,"test":[0.6706559196]},
-{"learn":[0.6604566326],"iteration":531,"passed_time":17.80870208,"remaining_time":49.14130574,"test":[0.6706515072]},
-{"learn":[0.6604430839],"iteration":532,"passed_time":17.82904188,"remaining_time":49.07167811,"test":[0.6706474616]},
-{"learn":[0.6604273738],"iteration":533,"passed_time":17.86246645,"remaining_time":49.03815696,"test":[0.6706424204]},
-{"learn":[0.6604048016],"iteration":534,"passed_time":17.90552779,"remaining_time":49.03102469,"test":[0.6706520008]},
-{"learn":[0.6603845173],"iteration":535,"passed_time":18.02843143,"remaining_time":49.24183511,"test":[0.6706448306]},
-{"learn":[0.6603669212],"iteration":536,"passed_time":18.07245966,"remaining_time":49.23651485,"test":[0.6706415789]},
-{"learn":[0.6603488983],"iteration":537,"passed_time":18.10631942,"remaining_time":49.20341819,"test":[0.6706305359]},
-{"learn":[0.6603176881],"iteration":538,"passed_time":18.13531438,"remaining_time":49.1571323,"test":[0.6706152774]},
-{"learn":[0.6602953862],"iteration":539,"passed_time":18.16575265,"remaining_time":49.11481272,"test":[0.670616585]},
-{"learn":[0.6602672025],"iteration":540,"passed_time":18.20025584,"remaining_time":49.08349958,"test":[0.6705963243]},
-{"learn":[0.6602568636],"iteration":541,"passed_time":18.22381751,"remaining_time":49.02274158,"test":[0.6706027368]},
-{"learn":[0.660235705],"iteration":542,"passed_time":18.25438575,"remaining_time":48.98092088,"test":[0.6706003522]},
-{"learn":[0.6602152295],"iteration":543,"passed_time":18.28070524,"remaining_time":48.9277699,"test":[0.6706044301]},
-{"learn":[0.6601897709],"iteration":544,"passed_time":18.30768805,"remaining_time":48.87648827,"test":[0.6706047241]},
-{"learn":[0.6601683731],"iteration":545,"passed_time":18.33807201,"remaining_time":48.83435294,"test":[0.6706038235]},
-{"learn":[0.6601472267],"iteration":546,"passed_time":18.36776304,"remaining_time":48.79041993,"test":[0.6706026913]},
-{"learn":[0.6601262337],"iteration":547,"passed_time":18.41134623,"remaining_time":48.78334803,"test":[0.6705845786]},
-{"learn":[0.6601119991],"iteration":548,"passed_time":18.44405381,"remaining_time":48.74739905,"test":[0.6705873967]},
-{"learn":[0.6600869973],"iteration":549,"passed_time":18.47010718,"remaining_time":48.69391893,"test":[0.6705755426]},
-{"learn":[0.6600667497],"iteration":550,"passed_time":18.5036553,"remaining_time":48.66024779,"test":[0.6705715731]},
-{"learn":[0.6600397508],"iteration":551,"passed_time":18.53164471,"remaining_time":48.61199556,"test":[0.6705757153]},
-{"learn":[0.660016863],"iteration":552,"passed_time":18.5577607,"remaining_time":48.55891452,"test":[0.6705516814]},
-{"learn":[0.6599933158],"iteration":553,"passed_time":18.58492994,"remaining_time":48.50867995,"test":[0.6705530864]},
-{"learn":[0.6599632649],"iteration":554,"passed_time":18.62562092,"remaining_time":48.49373376,"test":[0.6705552479]},
-{"learn":[0.6599446007],"iteration":555,"passed_time":18.65010209,"remaining_time":48.43659608,"test":[0.6705563336]},
-{"learn":[0.6599138126],"iteration":556,"passed_time":18.67796421,"remaining_time":48.38833458,"test":[0.6705718544]},
-{"learn":[0.6598965504],"iteration":557,"passed_time":18.70319381,"remaining_time":48.33334314,"test":[0.6705688384]},
-{"learn":[0.6598785723],"iteration":558,"passed_time":18.72995694,"remaining_time":48.28241136,"test":[0.6705641528]},
-{"learn":[0.659860838],"iteration":559,"passed_time":18.75657945,"remaining_time":48.23120429,"test":[0.6705628467]},
-{"learn":[0.6598408724],"iteration":560,"passed_time":18.78181322,"remaining_time":48.17652269,"test":[0.670558488]},
-{"learn":[0.6598244857],"iteration":561,"passed_time":18.80867415,"remaining_time":48.12610931,"test":[0.6705544404]},
-{"learn":[0.6598082469],"iteration":562,"passed_time":18.83488797,"remaining_time":48.0741279,"test":[0.6705617451]},
-{"learn":[0.6597851673],"iteration":563,"passed_time":18.86939449,"remaining_time":48.04335193,"test":[0.6705631717]},
-{"learn":[0.6597683521],"iteration":564,"passed_time":18.90235988,"remaining_time":48.00864854,"test":[0.6705636201]},
-{"learn":[0.6597479006],"iteration":565,"passed_time":18.93001053,"remaining_time":47.96048604,"test":[0.6705537522]},
-{"learn":[0.6597310938],"iteration":566,"passed_time":18.95858079,"remaining_time":47.91472006,"test":[0.670555083]},
-{"learn":[0.6597096581],"iteration":567,"passed_time":18.9833487,"remaining_time":47.85942842,"test":[0.6705524541]},
-{"learn":[0.6596862311],"iteration":568,"passed_time":19.0162481,"remaining_time":47.82469425,"test":[0.6705503132]},
-{"learn":[0.6596574779],"iteration":569,"passed_time":19.03781666,"remaining_time":47.76154004,"test":[0.6705354602]},
-{"learn":[0.6596385418],"iteration":570,"passed_time":19.0681355,"remaining_time":47.72043018,"test":[0.6705387012]},
-{"learn":[0.6596189903],"iteration":571,"passed_time":19.09073714,"remaining_time":47.66009201,"test":[0.6705411923]},
-{"learn":[0.65959275],"iteration":572,"passed_time":19.11146842,"remaining_time":47.59522765,"test":[0.6705390018]},
-{"learn":[0.6595730662],"iteration":573,"passed_time":19.141368,"remaining_time":47.55329403,"test":[0.6705354939]},
-{"learn":[0.6595566809],"iteration":574,"passed_time":19.16428373,"remaining_time":47.49409447,"test":[0.670531296]},
-{"learn":[0.6595365076],"iteration":575,"passed_time":19.19652276,"remaining_time":47.45807015,"test":[0.6705377163]},
-{"learn":[0.6595163446],"iteration":576,"passed_time":19.21727405,"remaining_time":47.39372785,"test":[0.6705248875]},
-{"learn":[0.6594816637],"iteration":577,"passed_time":19.24969594,"remaining_time":47.35824848,"test":[0.6705252902]},
-{"learn":[0.6594570142],"iteration":578,"passed_time":19.27445137,"remaining_time":47.30396442,"test":[0.6705181562]},
-{"learn":[0.6594353055],"iteration":579,"passed_time":19.29822455,"remaining_time":47.24737734,"test":[0.6705123446]},
-{"learn":[0.6594162362],"iteration":580,"passed_time":19.32403522,"remaining_time":47.19587948,"test":[0.6705128345]},
-{"learn":[0.659395036],"iteration":581,"passed_time":19.35739555,"remaining_time":47.16286408,"test":[0.6705173712]},
-{"learn":[0.6593798831],"iteration":582,"passed_time":19.39112791,"remaining_time":47.13075172,"test":[0.670541941]},
-{"learn":[0.6593556719],"iteration":583,"passed_time":19.42704318,"remaining_time":47.1039266,"test":[0.6705463243]},
-{"learn":[0.6593292627],"iteration":584,"passed_time":19.46022169,"remaining_time":47.07045077,"test":[0.6705513215]},
-{"learn":[0.6592976737],"iteration":585,"passed_time":19.48332075,"remaining_time":47.01265452,"test":[0.6705455889]},
-{"learn":[0.6592754841],"iteration":586,"passed_time":19.5115578,"remaining_time":46.9673444,"test":[0.6705408087]},
-{"learn":[0.6592510441],"iteration":587,"passed_time":19.54275193,"remaining_time":46.92919341,"test":[0.6705510193]},
-{"learn":[0.6592290326],"iteration":588,"passed_time":19.56411389,"remaining_time":46.86751222,"test":[0.6705456751]},
-{"learn":[0.6592097404],"iteration":589,"passed_time":19.59700884,"remaining_time":46.8335296,"test":[0.6705402427]},
-{"learn":[0.6591876204],"iteration":590,"passed_time":19.62169623,"remaining_time":46.77998306,"test":[0.6705443402]},
-{"learn":[0.6591705995],"iteration":591,"passed_time":19.64747626,"remaining_time":46.72913272,"test":[0.67054441]},
-{"learn":[0.6591456195],"iteration":592,"passed_time":19.67090184,"remaining_time":46.67278059,"test":[0.6705441955]},
-{"learn":[0.6591107122],"iteration":593,"passed_time":19.69910949,"remaining_time":46.62785848,"test":[0.6705319356]},
-{"learn":[0.6590819533],"iteration":594,"passed_time":19.72694709,"remaining_time":46.58211876,"test":[0.6705358843]},
-{"learn":[0.6590551327],"iteration":595,"passed_time":19.7530808,"remaining_time":46.53242523,"test":[0.6705334396]},
-{"learn":[0.6590373916],"iteration":596,"passed_time":19.77835609,"remaining_time":46.48079328,"test":[0.6705320462]},
-{"learn":[0.6590177149],"iteration":597,"passed_time":19.80378809,"remaining_time":46.4296169,"test":[0.6705332043]},
-{"learn":[0.6589946095],"iteration":598,"passed_time":19.83052585,"remaining_time":46.38158048,"test":[0.6705328363]},
-{"learn":[0.6589697628],"iteration":599,"passed_time":19.8579153,"remaining_time":46.33513569,"test":[0.6705315638]},
-{"learn":[0.6589442269],"iteration":600,"passed_time":19.89600309,"remaining_time":46.31365777,"test":[0.6705274435]},
-{"learn":[0.6589182437],"iteration":601,"passed_time":19.92518872,"remaining_time":46.27145155,"test":[0.670509808]},
-{"learn":[0.6588837179],"iteration":602,"passed_time":19.95754179,"remaining_time":46.23662666,"test":[0.6705077789]},
-{"learn":[0.6588674101],"iteration":603,"passed_time":19.99116426,"remaining_time":46.20474388,"test":[0.6705212132]},
-{"learn":[0.6588406916],"iteration":604,"passed_time":20.01900069,"remaining_time":46.15951398,"test":[0.6705098442]},
-{"learn":[0.6588149945],"iteration":605,"passed_time":20.04735837,"remaining_time":46.11554053,"test":[0.6705061509]},
-{"learn":[0.6587866031],"iteration":606,"passed_time":20.07232044,"remaining_time":46.06382599,"test":[0.6705003071]},
-{"learn":[0.6587636648],"iteration":607,"passed_time":20.09871086,"remaining_time":46.01546959,"test":[0.6705045031]},
-{"learn":[0.6587502469],"iteration":608,"passed_time":20.12348304,"remaining_time":45.96348917,"test":[0.6705083194]},
-{"learn":[0.6587292784],"iteration":609,"passed_time":20.14920752,"remaining_time":45.91376797,"test":[0.6705329997]},
-{"learn":[0.6587104112],"iteration":610,"passed_time":20.17662353,"remaining_time":45.86797068,"test":[0.6705269987]},
-{"learn":[0.6586953782],"iteration":611,"passed_time":20.20202219,"remaining_time":45.81765818,"test":[0.6705315607]},
-{"learn":[0.6586641191],"iteration":612,"passed_time":20.23050051,"remaining_time":45.77439512,"test":[0.6705142835]},
-{"learn":[0.6586450136],"iteration":613,"passed_time":20.25381994,"remaining_time":45.71953492,"test":[0.6705165015]},
-{"learn":[0.6586136263],"iteration":614,"passed_time":20.28518384,"remaining_time":45.68289369,"test":[0.6705001061]},
-{"learn":[0.6585862768],"iteration":615,"passed_time":20.3078175,"remaining_time":45.62665489,"test":[0.6705013916]},
-{"learn":[0.6585585235],"iteration":616,"passed_time":20.33878033,"remaining_time":45.5891948,"test":[0.6705037253]},
-{"learn":[0.6585371631],"iteration":617,"passed_time":20.36122842,"remaining_time":45.53271469,"test":[0.67049647]},
-{"learn":[0.6585092632],"iteration":618,"passed_time":20.3943397,"remaining_time":45.50013429,"test":[0.6705005632]},
-{"learn":[0.6584914317],"iteration":619,"passed_time":20.42384285,"remaining_time":45.45952119,"test":[0.6704957943]},
-{"learn":[0.6584662432],"iteration":620,"passed_time":20.45411533,"remaining_time":45.42065225,"test":[0.6704955333]},
-{"learn":[0.6584454668],"iteration":621,"passed_time":20.488223,"remaining_time":45.39030754,"test":[0.6704961207]},
-{"learn":[0.6584249408],"iteration":622,"passed_time":20.51043528,"remaining_time":45.33365872,"test":[0.6704921459]},
-{"learn":[0.6583931228],"iteration":623,"passed_time":20.54384208,"remaining_time":45.30180561,"test":[0.6704751713]},
-{"learn":[0.6583660767],"iteration":624,"passed_time":20.56912557,"remaining_time":45.25207624,"test":[0.6704753101]},
-{"learn":[0.658354264],"iteration":625,"passed_time":20.59414123,"remaining_time":45.20183714,"test":[0.6704620888]},
-{"learn":[0.6583253625],"iteration":626,"passed_time":20.61901142,"remaining_time":45.15135993,"test":[0.6704604282]},
-{"learn":[0.6582968632],"iteration":627,"passed_time":20.6468542,"remaining_time":45.10745855,"test":[0.6704663192]},
-{"learn":[0.6582687399],"iteration":628,"passed_time":20.67583093,"remaining_time":45.06607981,"test":[0.6704680085]},
-{"learn":[0.658242535],"iteration":629,"passed_time":20.7010198,"remaining_time":45.01650336,"test":[0.670453228]},
-{"learn":[0.6582199874],"iteration":630,"passed_time":20.72783977,"remaining_time":44.97054302,"test":[0.6704577785]},
-{"learn":[0.6581918101],"iteration":631,"passed_time":20.75222724,"remaining_time":44.91937795,"test":[0.67046675]},
-{"learn":[0.6581735218],"iteration":632,"passed_time":20.78264004,"remaining_time":44.88130954,"test":[0.6704731863]},
-{"learn":[0.6581445869],"iteration":633,"passed_time":20.80459182,"remaining_time":44.82503538,"test":[0.6704811116]},
-{"learn":[0.6581202427],"iteration":634,"passed_time":20.83717209,"remaining_time":44.79171637,"test":[0.6704839644]},
-{"learn":[0.6580977862],"iteration":635,"passed_time":20.86231353,"remaining_time":44.74244599,"test":[0.6704854798]},
-{"learn":[0.6580724179],"iteration":636,"passed_time":20.89269601,"remaining_time":44.70446572,"test":[0.6704835837]},
-{"learn":[0.6580426322],"iteration":637,"passed_time":20.93117347,"remaining_time":44.68379039,"test":[0.6704736198]},
-{"learn":[0.6580111256],"iteration":638,"passed_time":20.96066949,"remaining_time":44.64392985,"test":[0.6704640242]},
-{"learn":[0.6579834747],"iteration":639,"passed_time":20.9941179,"remaining_time":44.61250055,"test":[0.670465663]},
-{"learn":[0.6579541367],"iteration":640,"passed_time":21.0224519,"remaining_time":44.57022174,"test":[0.6704646829]},
-{"learn":[0.6579254503],"iteration":641,"passed_time":21.0522529,"remaining_time":44.53108946,"test":[0.6704600961]},
-{"learn":[0.657898555],"iteration":642,"passed_time":21.08260618,"remaining_time":44.49315178,"test":[0.6704643207]},
-{"learn":[0.6578676875],"iteration":643,"passed_time":21.10716702,"remaining_time":44.44304112,"test":[0.6704600533]},
-{"learn":[0.6578324163],"iteration":644,"passed_time":21.13594828,"remaining_time":44.40187584,"test":[0.6704614691]},
-{"learn":[0.6578062223],"iteration":645,"passed_time":21.1601277,"remaining_time":44.35110357,"test":[0.6704728212]},
-{"learn":[0.6577760631],"iteration":646,"passed_time":21.18552999,"remaining_time":44.30297075,"test":[0.6704758731]},
-{"learn":[0.6577483474],"iteration":647,"passed_time":21.21048648,"remaining_time":44.25397797,"test":[0.6704833026]},
-{"learn":[0.6577249642],"iteration":648,"passed_time":21.23686209,"remaining_time":44.20801337,"test":[0.6704767664]},
-{"learn":[0.6576974966],"iteration":649,"passed_time":21.26287585,"remaining_time":44.16135753,"test":[0.6704702727]},
-{"learn":[0.657675114],"iteration":650,"passed_time":21.28806218,"remaining_time":44.11305051,"test":[0.6704671372]},
-{"learn":[0.6576447891],"iteration":651,"passed_time":21.31506267,"remaining_time":44.06856515,"test":[0.6704699936]},
-{"learn":[0.6576102356],"iteration":652,"passed_time":21.3435081,"remaining_time":44.02711394,"test":[0.6704587989]},
-{"learn":[0.6575793887],"iteration":653,"passed_time":21.37776713,"remaining_time":43.99766753,"test":[0.6704637668]},
-{"learn":[0.6575543309],"iteration":654,"passed_time":21.40301154,"remaining_time":43.94969545,"test":[0.6704653717]},
-{"learn":[0.6575340787],"iteration":655,"passed_time":21.44023109,"remaining_time":43.92632711,"test":[0.6704598273]},
-{"learn":[0.6575061464],"iteration":656,"passed_time":21.4778965,"remaining_time":43.903828,"test":[0.6704522865]},
-{"learn":[0.657476113],"iteration":657,"passed_time":21.50245582,"remaining_time":43.85455275,"test":[0.6704558586]},
-{"learn":[0.6574447014],"iteration":658,"passed_time":21.53379663,"remaining_time":43.81915217,"test":[0.6704466331]},
-{"learn":[0.6574247361],"iteration":659,"passed_time":21.55955041,"remaining_time":43.77242053,"test":[0.6704405886]},
-{"learn":[0.6574034983],"iteration":660,"passed_time":21.58626671,"remaining_time":43.72770215,"test":[0.6704463767]},
-{"learn":[0.6573783832],"iteration":661,"passed_time":21.61183918,"remaining_time":43.68072633,"test":[0.6704475216]},
-{"learn":[0.657357694],"iteration":662,"passed_time":21.6373217,"remaining_time":43.63363366,"test":[0.6704572386]},
-{"learn":[0.6573411592],"iteration":663,"passed_time":21.66283476,"remaining_time":43.58666753,"test":[0.6704658153]},
-{"learn":[0.6573118559],"iteration":664,"passed_time":21.68841321,"remaining_time":43.5398972,"test":[0.6704600945]},
-{"learn":[0.6572819076],"iteration":665,"passed_time":21.71420973,"remaining_time":43.4936273,"test":[0.6704561998]},
-{"learn":[0.6572430097],"iteration":666,"passed_time":21.74213421,"remaining_time":43.45167151,"test":[0.6704535154]},
-{"learn":[0.6572160391],"iteration":667,"passed_time":21.77174463,"remaining_time":43.41311953,"test":[0.6704413781]},
-{"learn":[0.6571931413],"iteration":668,"passed_time":21.81895309,"remaining_time":43.40960622,"test":[0.6704450013]},
-{"learn":[0.6571737099],"iteration":669,"passed_time":21.84627583,"remaining_time":43.36648784,"test":[0.6704422199]},
-{"learn":[0.6571532872],"iteration":670,"passed_time":21.88834724,"remaining_time":43.35262814,"test":[0.67044342]},
-{"learn":[0.6571208939],"iteration":671,"passed_time":21.93403139,"remaining_time":43.34582395,"test":[0.6704415341]},
-{"learn":[0.6570887673],"iteration":672,"passed_time":21.9714274,"remaining_time":43.32256191,"test":[0.6704439539]},
-{"learn":[0.6570633692],"iteration":673,"passed_time":22.01942449,"remaining_time":43.32011406,"test":[0.6704498197]},
-{"learn":[0.6570454361],"iteration":674,"passed_time":22.05319867,"remaining_time":43.2896122,"test":[0.6704452194]},
-{"learn":[0.6570231031],"iteration":675,"passed_time":22.09079747,"remaining_time":43.26659149,"test":[0.6704366524]},
-{"learn":[0.6570052089],"iteration":676,"passed_time":22.14192346,"remaining_time":43.26996269,"test":[0.6704427124]},
-{"learn":[0.6569855794],"iteration":677,"passed_time":22.17624471,"remaining_time":43.24040635,"test":[0.6704395579]},
-{"learn":[0.6569579709],"iteration":678,"passed_time":22.213192,"remaining_time":43.21594497,"test":[0.6704401246]},
-{"learn":[0.6569333354],"iteration":679,"passed_time":22.23966403,"remaining_time":43.17111253,"test":[0.6704415621]},
-{"learn":[0.6569069617],"iteration":680,"passed_time":22.27051241,"remaining_time":43.13481039,"test":[0.6704341343]},
-{"learn":[0.6568931857],"iteration":681,"passed_time":22.29625075,"remaining_time":43.08864881,"test":[0.6704369615]},
-{"learn":[0.6568734532],"iteration":682,"passed_time":22.32160622,"remaining_time":43.04180877,"test":[0.6704357425]},
-{"learn":[0.6568435196],"iteration":683,"passed_time":22.35059872,"remaining_time":43.00202911,"test":[0.6704294622]},
-{"learn":[0.6568108038],"iteration":684,"passed_time":22.37956576,"remaining_time":42.96223208,"test":[0.6704289794]},
-{"learn":[0.6567811374],"iteration":685,"passed_time":22.41993338,"remaining_time":42.94430389,"test":[0.6704272409]},
-{"learn":[0.6567467284],"iteration":686,"passed_time":22.45285267,"remaining_time":42.91207504,"test":[0.6704101162]},
-{"learn":[0.6567172734],"iteration":687,"passed_time":22.4848431,"remaining_time":42.8780729,"test":[0.6704069439]},
-{"learn":[0.6566967606],"iteration":688,"passed_time":22.51193834,"remaining_time":42.83476221,"test":[0.6704100747]},
-{"learn":[0.6566720128],"iteration":689,"passed_time":22.53798671,"remaining_time":42.78951101,"test":[0.6704122261]},
-{"learn":[0.6566441608],"iteration":690,"passed_time":22.57108439,"remaining_time":42.75766928,"test":[0.6704137826]},
-{"learn":[0.6566172287],"iteration":691,"passed_time":22.59836588,"remaining_time":42.7148303,"test":[0.6704207952]},
-{"learn":[0.6565952549],"iteration":692,"passed_time":22.62447507,"remaining_time":42.66982528,"test":[0.6704154834]},
-{"learn":[0.6565702687],"iteration":693,"passed_time":22.65349415,"remaining_time":42.63035067,"test":[0.6704253514]},
-{"learn":[0.6565392213],"iteration":694,"passed_time":22.68028991,"remaining_time":42.58673141,"test":[0.6704155636]},
-{"learn":[0.6565157938],"iteration":695,"passed_time":22.70844406,"remaining_time":42.54570555,"test":[0.6704141298]},
-{"learn":[0.6564902789],"iteration":696,"passed_time":22.73944116,"remaining_time":42.51003133,"test":[0.6704207635]},
-{"learn":[0.6564644734],"iteration":697,"passed_time":22.7613976,"remaining_time":42.45750671,"test":[0.6704268341]},
-{"learn":[0.6564349549],"iteration":698,"passed_time":22.79216825,"remaining_time":42.42147482,"test":[0.6704243126]},
-{"learn":[0.6564046572],"iteration":699,"passed_time":22.8167121,"remaining_time":42.37389389,"test":[0.6704235165]},
-{"learn":[0.6563744107],"iteration":700,"passed_time":22.84507296,"remaining_time":42.33345189,"test":[0.6704257736]},
-{"learn":[0.6563525063],"iteration":701,"passed_time":22.87088832,"remaining_time":42.28833766,"test":[0.6704247758]},
-{"learn":[0.6563189867],"iteration":702,"passed_time":22.90238907,"remaining_time":42.25376759,"test":[0.6704331799]},
-{"learn":[0.6562939062],"iteration":703,"passed_time":22.94246813,"remaining_time":42.23499815,"test":[0.6704252722]},
-{"learn":[0.6562739297],"iteration":704,"passed_time":22.97441688,"remaining_time":42.20123385,"test":[0.6704146644]},
-{"learn":[0.656256438],"iteration":705,"passed_time":23.00262167,"remaining_time":42.16061253,"test":[0.6704164122]},
-{"learn":[0.6562366475],"iteration":706,"passed_time":23.033437,"remaining_time":42.12480062,"test":[0.6704118954]},
-{"learn":[0.6562073096],"iteration":707,"passed_time":23.0545813,"remaining_time":42.07135458,"test":[0.6704043129]},
-{"learn":[0.6561864222],"iteration":708,"passed_time":23.08699831,"remaining_time":42.03852584,"test":[0.6703978198]},
-{"learn":[0.6561578826],"iteration":709,"passed_time":23.11590694,"remaining_time":41.99932387,"test":[0.6703935976]},
-{"learn":[0.6561208567],"iteration":710,"passed_time":23.14362702,"remaining_time":41.9579961,"test":[0.6703839683]},
-{"learn":[0.6560924703],"iteration":711,"passed_time":23.16985155,"remaining_time":41.91400112,"test":[0.6703843723]},
-{"learn":[0.6560656907],"iteration":712,"passed_time":23.19510285,"remaining_time":41.86829925,"test":[0.6703879502]},
-{"learn":[0.6560362588],"iteration":713,"passed_time":23.23034771,"remaining_time":41.84065429,"test":[0.6703895978]},
-{"learn":[0.6560124527],"iteration":714,"passed_time":23.25923754,"remaining_time":41.80156678,"test":[0.6703894359]},
-{"learn":[0.6559875055],"iteration":715,"passed_time":23.28703452,"remaining_time":41.76054794,"test":[0.6703928777]},
-{"learn":[0.6559547281],"iteration":716,"passed_time":23.31161175,"remaining_time":41.71380457,"test":[0.6703933128]},
-{"learn":[0.6559230866],"iteration":717,"passed_time":23.34170355,"remaining_time":41.67696929,"test":[0.6703844355]},
-{"learn":[0.6558924823],"iteration":718,"passed_time":23.37263658,"remaining_time":41.64165155,"test":[0.6703825151]},
-{"learn":[0.6558676469],"iteration":719,"passed_time":23.40571088,"remaining_time":41.61015268,"test":[0.6703983542]},
-{"learn":[0.6558459277],"iteration":720,"passed_time":23.4389719,"remaining_time":41.57898067,"test":[0.670399556]},
-{"learn":[0.6558149638],"iteration":721,"passed_time":23.48304084,"remaining_time":41.56693379,"test":[0.6703931808]},
-{"learn":[0.6557812248],"iteration":722,"passed_time":23.50734531,"remaining_time":41.5198893,"test":[0.6703886918]},
-{"learn":[0.6557546502],"iteration":723,"passed_time":23.54055835,"remaining_time":41.48860836,"test":[0.6703847574]},
-{"learn":[0.6557274948],"iteration":724,"passed_time":23.56652491,"remaining_time":41.44457829,"test":[0.6703885941]},
-{"learn":[0.6557044723],"iteration":725,"passed_time":23.59580183,"remaining_time":41.40640708,"test":[0.6703788615]},
-{"learn":[0.6556751811],"iteration":726,"passed_time":23.62334313,"remaining_time":41.36522119,"test":[0.6703799906]},
-{"learn":[0.6556539158],"iteration":727,"passed_time":23.64879831,"remaining_time":41.32042782,"test":[0.6703774518]},
-{"learn":[0.6556182915],"iteration":728,"passed_time":23.67755213,"remaining_time":41.28143862,"test":[0.6703783496]},
-{"learn":[0.6555977079],"iteration":729,"passed_time":23.70012944,"remaining_time":41.23173204,"test":[0.6703648854]},
-{"learn":[0.6555667903],"iteration":730,"passed_time":23.72866102,"remaining_time":41.19243615,"test":[0.6703716654]},
-{"learn":[0.6555394075],"iteration":731,"passed_time":23.75226732,"remaining_time":41.14463793,"test":[0.6703550938]},
-{"learn":[0.6555122742],"iteration":732,"passed_time":23.7844108,"remaining_time":41.11166233,"test":[0.6703467057]},
-{"learn":[0.6554814941],"iteration":733,"passed_time":23.80747563,"remaining_time":41.06303017,"test":[0.6703484503]},
-{"learn":[0.6554517373],"iteration":734,"passed_time":23.84023587,"remaining_time":41.03115425,"test":[0.6703549183]},
-{"learn":[0.655429552],"iteration":735,"passed_time":23.87042124,"remaining_time":40.99485387,"test":[0.6703501504]},
-{"learn":[0.655396579],"iteration":736,"passed_time":23.9087808,"remaining_time":40.97257823,"test":[0.6703672622]},
-{"learn":[0.6553735864],"iteration":737,"passed_time":23.94161529,"remaining_time":40.94081097,"test":[0.6703560249]},
-{"learn":[0.6553472597],"iteration":738,"passed_time":23.97478791,"remaining_time":40.90961779,"test":[0.6703547155]},
-{"learn":[0.6553252832],"iteration":739,"passed_time":24.00628859,"remaining_time":40.87557247,"test":[0.6703593236]},
-{"learn":[0.6552971659],"iteration":740,"passed_time":24.03623034,"remaining_time":40.83888528,"test":[0.6703606827]},
-{"learn":[0.6552763852],"iteration":741,"passed_time":24.06404686,"remaining_time":40.79861313,"test":[0.6703511404]},
-{"learn":[0.6552488203],"iteration":742,"passed_time":24.09270947,"remaining_time":40.75980593,"test":[0.6703431646]},
-{"learn":[0.65521229],"iteration":743,"passed_time":24.12724624,"remaining_time":40.73094258,"test":[0.6703475116]},
-{"learn":[0.6551949744],"iteration":744,"passed_time":24.15397955,"remaining_time":40.68891857,"test":[0.6703483634]},
-{"learn":[0.6551673797],"iteration":745,"passed_time":24.17955779,"remaining_time":40.64499392,"test":[0.6703475713]},
-{"learn":[0.6551421856],"iteration":746,"passed_time":24.20715317,"remaining_time":40.60450191,"test":[0.670360457]},
-{"learn":[0.6551255516],"iteration":747,"passed_time":24.23336836,"remaining_time":40.5617342,"test":[0.6703664352]},
-{"learn":[0.6551019608],"iteration":748,"passed_time":24.2614437,"remaining_time":40.52211759,"test":[0.6703617612]},
-{"learn":[0.6550758728],"iteration":749,"passed_time":24.29512083,"remaining_time":40.49186805,"test":[0.6703669926]},
-{"learn":[0.655051966],"iteration":750,"passed_time":24.31839238,"remaining_time":40.44430371,"test":[0.6703670837]},
-{"learn":[0.6550351058],"iteration":751,"passed_time":24.34977118,"remaining_time":40.41025856,"test":[0.6703706628]},
-{"learn":[0.6549998756],"iteration":752,"passed_time":24.3762114,"remaining_time":40.36804198,"test":[0.670369618]},
-{"learn":[0.6549721212],"iteration":753,"passed_time":24.40831154,"remaining_time":40.3352204,"test":[0.6703692351]},
-{"learn":[0.6549401744],"iteration":754,"passed_time":24.44267281,"remaining_time":40.30612934,"test":[0.6703624433]},
-{"learn":[0.6549207325],"iteration":755,"passed_time":24.47460721,"remaining_time":40.27303091,"test":[0.6703686285]},
-{"learn":[0.6548900891],"iteration":756,"passed_time":24.50826603,"remaining_time":40.24276708,"test":[0.6703598432]},
-{"learn":[0.6548682731],"iteration":757,"passed_time":24.54826542,"remaining_time":40.22288345,"test":[0.6703618766]},
-{"learn":[0.6548418938],"iteration":758,"passed_time":24.57546587,"remaining_time":40.18201996,"test":[0.6703694148]},
-{"learn":[0.6548234717],"iteration":759,"passed_time":24.60502723,"remaining_time":40.14504442,"test":[0.6703683652]},
-{"learn":[0.6547996833],"iteration":760,"passed_time":24.63261096,"remaining_time":40.10486856,"test":[0.6703604855]},
-{"learn":[0.6547726174],"iteration":761,"passed_time":24.66001655,"remaining_time":40.06443634,"test":[0.6703758987]},
-{"learn":[0.6547509314],"iteration":762,"passed_time":24.68929907,"remaining_time":40.02708119,"test":[0.6703773302]},
-{"learn":[0.6547168175],"iteration":763,"passed_time":24.71425118,"remaining_time":39.98274144,"test":[0.6703641028]},
-{"learn":[0.6546907846],"iteration":764,"passed_time":24.74589169,"remaining_time":39.94924999,"test":[0.6703649602]},
-{"learn":[0.6546671611],"iteration":765,"passed_time":24.76625006,"remaining_time":39.89758822,"test":[0.6703567811]},
-{"learn":[0.6546475893],"iteration":766,"passed_time":24.79734832,"remaining_time":39.86327312,"test":[0.6703544688]},
-{"learn":[0.6546206223],"iteration":767,"passed_time":24.82531049,"remaining_time":39.82393558,"test":[0.6703611821]},
-{"learn":[0.6545874193],"iteration":768,"passed_time":24.85435247,"remaining_time":39.78635616,"test":[0.6703527821]},
-{"learn":[0.6545620629],"iteration":769,"passed_time":24.88095966,"remaining_time":39.74490958,"test":[0.6703523616]},
-{"learn":[0.6545346297],"iteration":770,"passed_time":24.90935211,"remaining_time":39.70634726,"test":[0.6703616298]},
-{"learn":[0.6545172316],"iteration":771,"passed_time":24.94098876,"remaining_time":39.67297175,"test":[0.6703603551]},
-{"learn":[0.6544943049],"iteration":772,"passed_time":24.97035098,"remaining_time":39.6359905,"test":[0.6703675655]},
-{"learn":[0.6544632323],"iteration":773,"passed_time":25.00434422,"remaining_time":39.60636436,"test":[0.6703582411]},
-{"learn":[0.6544384097],"iteration":774,"passed_time":25.03067441,"remaining_time":39.56461439,"test":[0.6703581437]},
-{"learn":[0.6544084745],"iteration":775,"passed_time":25.05692652,"remaining_time":39.522781,"test":[0.6703551885]},
-{"learn":[0.6543765257],"iteration":776,"passed_time":25.08660163,"remaining_time":39.48637554,"test":[0.6703608491]},
-{"learn":[0.6543536123],"iteration":777,"passed_time":25.10764591,"remaining_time":39.43643098,"test":[0.6703674554]},
-{"learn":[0.6543303593],"iteration":778,"passed_time":25.13940138,"remaining_time":39.40334928,"test":[0.6703679619]},
-{"learn":[0.6543005831],"iteration":779,"passed_time":25.15916899,"remaining_time":39.35152074,"test":[0.6703701757]},
-{"learn":[0.6542678123],"iteration":780,"passed_time":25.18841105,"remaining_time":39.31456219,"test":[0.6703603462]},
-{"learn":[0.6542439303],"iteration":781,"passed_time":25.21444083,"remaining_time":39.27262012,"test":[0.670359801]},
-{"learn":[0.6542100401],"iteration":782,"passed_time":25.24017824,"remaining_time":39.23026426,"test":[0.6703523669]},
-{"learn":[0.6541836178],"iteration":783,"passed_time":25.2660091,"remaining_time":39.18809574,"test":[0.6703365674]},
-{"learn":[0.654158129],"iteration":784,"passed_time":25.28891553,"remaining_time":39.1414425,"test":[0.6703486118]},
-{"learn":[0.6541343464],"iteration":785,"passed_time":25.31589904,"remaining_time":39.10114686,"test":[0.6703450011]},
-{"learn":[0.6541092921],"iteration":786,"passed_time":25.34123581,"remaining_time":39.05834694,"test":[0.6703473135]},
-{"learn":[0.6540812254],"iteration":787,"passed_time":25.36728606,"remaining_time":39.01668871,"test":[0.670350998]},
-{"learn":[0.654060259],"iteration":788,"passed_time":25.39177931,"remaining_time":38.97268028,"test":[0.6703417767]},
-{"learn":[0.6540467253],"iteration":789,"passed_time":25.41712461,"remaining_time":38.9300263,"test":[0.6703349821]},
-{"learn":[0.6540306837],"iteration":790,"passed_time":25.44804125,"remaining_time":38.89593157,"test":[0.6703457717]},
-{"learn":[0.6540103667],"iteration":791,"passed_time":25.48249341,"remaining_time":38.86723743,"test":[0.6703506266]},
-{"learn":[0.6539821302],"iteration":792,"passed_time":25.51450657,"remaining_time":38.83481643,"test":[0.6703596395]},
-{"learn":[0.6539577914],"iteration":793,"passed_time":25.54216564,"remaining_time":38.79578307,"test":[0.6703799895]},
-{"learn":[0.653923724],"iteration":794,"passed_time":25.56982738,"remaining_time":38.75678238,"test":[0.6703687687]},
-{"learn":[0.6539086888],"iteration":795,"passed_time":25.59539769,"remaining_time":38.71464675,"test":[0.6703780675]},
-{"learn":[0.6538798424],"iteration":796,"passed_time":25.61874122,"remaining_time":38.66919157,"test":[0.670374835]},
-{"learn":[0.6538566996],"iteration":797,"passed_time":25.64394874,"remaining_time":38.62659947,"test":[0.6703831387]},
-{"learn":[0.6538290752],"iteration":798,"passed_time":25.66776244,"remaining_time":38.58195581,"test":[0.670377656]},
-{"learn":[0.6538051255],"iteration":799,"passed_time":25.69593415,"remaining_time":38.54390122,"test":[0.6703689741]},
-{"learn":[0.6537917354],"iteration":800,"passed_time":25.71651353,"remaining_time":38.49450652,"test":[0.6703709756]},
-{"learn":[0.6537684302],"iteration":801,"passed_time":25.74304126,"remaining_time":38.45406912,"test":[0.6703737517]},
-{"learn":[0.6537402991],"iteration":802,"passed_time":25.77084871,"remaining_time":38.41557398,"test":[0.6703818964]},
-{"learn":[0.6537165427],"iteration":803,"passed_time":25.79028824,"remaining_time":38.36465763,"test":[0.6703812173]},
-{"learn":[0.6536853601],"iteration":804,"passed_time":25.82203653,"remaining_time":38.3320915,"test":[0.6703960068]},
-{"learn":[0.6536681479],"iteration":805,"passed_time":25.84395064,"remaining_time":38.28495914,"test":[0.6703976729]},
-{"learn":[0.6536409101],"iteration":806,"passed_time":25.87390688,"remaining_time":38.24977808,"test":[0.6704024604]},
-{"learn":[0.6536120189],"iteration":807,"passed_time":25.89606204,"remaining_time":38.20310143,"test":[0.6704085008]},
-{"learn":[0.6535912493],"iteration":808,"passed_time":25.92585483,"remaining_time":38.16772942,"test":[0.6704076633]},
-{"learn":[0.6535617421],"iteration":809,"passed_time":25.95539059,"remaining_time":38.13199358,"test":[0.6704111719]},
-{"learn":[0.6535315174],"iteration":810,"passed_time":25.98822968,"remaining_time":38.10111601,"test":[0.6704220803]},
-{"learn":[0.6534972927],"iteration":811,"passed_time":26.02835773,"remaining_time":38.08089777,"test":[0.6704265011]},
-{"learn":[0.6534818476],"iteration":812,"passed_time":26.0558565,"remaining_time":38.04219146,"test":[0.6704251162]},
-{"learn":[0.6534498323],"iteration":813,"passed_time":26.08151817,"remaining_time":38.00083606,"test":[0.6704375472]},
-{"learn":[0.6534305025],"iteration":814,"passed_time":26.10848988,"remaining_time":37.96142393,"test":[0.6704319336]},
-{"learn":[0.6534081059],"iteration":815,"passed_time":26.13143346,"remaining_time":37.91619757,"test":[0.670437614]},
-{"learn":[0.6533765804],"iteration":816,"passed_time":26.15923661,"remaining_time":37.87806231,"test":[0.6704554331]},
-{"learn":[0.6533441549],"iteration":817,"passed_time":26.18805523,"remaining_time":37.84141966,"test":[0.6704603317]},
-{"learn":[0.6533053405],"iteration":818,"passed_time":26.2140726,"remaining_time":37.8007567,"test":[0.6704548042]},
-{"learn":[0.6532838469],"iteration":819,"passed_time":26.24289367,"remaining_time":37.76416405,"test":[0.6704502654]},
-{"learn":[0.6532604302],"iteration":820,"passed_time":26.27260776,"remaining_time":37.72887277,"test":[0.6704512072]},
-{"learn":[0.6532364412],"iteration":821,"passed_time":26.29880394,"remaining_time":37.68855358,"test":[0.6704433481]},
-{"learn":[0.6532100089],"iteration":822,"passed_time":26.32785215,"remaining_time":37.65234749,"test":[0.6704095112]},
-{"learn":[0.6531782515],"iteration":823,"passed_time":26.35925682,"remaining_time":37.61952188,"test":[0.6704086019]},
-{"learn":[0.6531449701],"iteration":824,"passed_time":26.38596096,"remaining_time":37.580005,"test":[0.6703987131]},
-{"learn":[0.653115452],"iteration":825,"passed_time":26.40854839,"remaining_time":37.53466805,"test":[0.6704019708]},
-{"learn":[0.6530787602],"iteration":826,"passed_time":26.44419918,"remaining_time":37.50791492,"test":[0.6704046556]},
-{"learn":[0.653052397],"iteration":827,"passed_time":26.47784276,"remaining_time":37.47829917,"test":[0.6704091961]},
-{"learn":[0.6530313579],"iteration":828,"passed_time":26.51701028,"remaining_time":37.45647652,"test":[0.6704103204]},
-{"learn":[0.6530010363],"iteration":829,"passed_time":26.53963123,"remaining_time":37.41128739,"test":[0.6704074257]},
-{"learn":[0.6529752146],"iteration":830,"passed_time":26.57362226,"remaining_time":37.38214732,"test":[0.6704115335]},
-{"learn":[0.652954801],"iteration":831,"passed_time":26.59767057,"remaining_time":37.33903754,"test":[0.6704041275]},
-{"learn":[0.6529330351],"iteration":832,"passed_time":26.62378941,"remaining_time":37.29887425,"test":[0.6704004556]},
-{"learn":[0.6528993709],"iteration":833,"passed_time":26.65024746,"remaining_time":37.25921887,"test":[0.6704037097]},
-{"learn":[0.6528665883],"iteration":834,"passed_time":26.67774911,"remaining_time":37.22105115,"test":[0.6704035477]},
-{"learn":[0.6528413041],"iteration":835,"passed_time":26.70473813,"remaining_time":37.1821952,"test":[0.6704025281]},
-{"learn":[0.6528217161],"iteration":836,"passed_time":26.72833235,"remaining_time":37.13865056,"test":[0.6704024549]},
-{"learn":[0.6527978782],"iteration":837,"passed_time":26.76384162,"remaining_time":37.11167537,"test":[0.670405721]},
-{"learn":[0.6527789461],"iteration":838,"passed_time":26.79137369,"remaining_time":37.07364106,"test":[0.6703983189]},
-{"learn":[0.6527432001],"iteration":839,"passed_time":26.82295602,"remaining_time":37.04122498,"test":[0.6704035256]},
-{"learn":[0.6527139767],"iteration":840,"passed_time":26.87217031,"remaining_time":37.03310985,"test":[0.6704047613]},
-{"learn":[0.6526857244],"iteration":841,"passed_time":26.92488006,"remaining_time":37.0297044,"test":[0.6704139617]},
-{"learn":[0.652657086],"iteration":842,"passed_time":26.98258041,"remaining_time":37.03303147,"test":[0.6704066193]},
-{"learn":[0.6526355016],"iteration":843,"passed_time":27.05424841,"remaining_time":37.05534497,"test":[0.670402892]},
-{"learn":[0.6526054936],"iteration":844,"passed_time":27.09765154,"remaining_time":37.03880181,"test":[0.6704081961]},
-{"learn":[0.6525793707],"iteration":845,"passed_time":27.12038959,"remaining_time":36.99400661,"test":[0.6704029862]},
-{"learn":[0.6525584692],"iteration":846,"passed_time":27.14691224,"remaining_time":36.95441537,"test":[0.6704014281]},
-{"learn":[0.6525279747],"iteration":847,"passed_time":27.18096334,"remaining_time":36.92508227,"test":[0.6704036115]},
-{"learn":[0.6525038765],"iteration":848,"passed_time":27.20686017,"remaining_time":36.88468322,"test":[0.6704016777]},
-{"learn":[0.6524849104],"iteration":849,"passed_time":27.23465701,"remaining_time":36.8468889,"test":[0.6704085392]},
-{"learn":[0.6524610603],"iteration":850,"passed_time":27.26094834,"remaining_time":36.80708536,"test":[0.6704042952]},
-{"learn":[0.6524357337],"iteration":851,"passed_time":27.28945577,"remaining_time":36.77029957,"test":[0.670394789]},
-{"learn":[0.6524082286],"iteration":852,"passed_time":27.31865398,"remaining_time":36.73446203,"test":[0.6703885644]},
-{"learn":[0.65238051],"iteration":853,"passed_time":27.34791322,"remaining_time":36.69872195,"test":[0.6703946813]},
-{"learn":[0.6523557826],"iteration":854,"passed_time":27.3865535,"remaining_time":36.67555995,"test":[0.6704042137]},
-{"learn":[0.6523391233],"iteration":855,"passed_time":27.41370907,"remaining_time":36.63701306,"test":[0.6704077517]},
-{"learn":[0.652325347],"iteration":856,"passed_time":27.43905921,"remaining_time":36.5960848,"test":[0.6704118698]},
-{"learn":[0.6522924958],"iteration":857,"passed_time":27.47159295,"remaining_time":36.56475425,"test":[0.6704114259]},
-{"learn":[0.6522623584],"iteration":858,"passed_time":27.50124299,"remaining_time":36.52959052,"test":[0.6704157567]},
-{"learn":[0.6522343891],"iteration":859,"passed_time":27.53509105,"remaining_time":36.50000442,"test":[0.6703837005]},
-{"learn":[0.6522094424],"iteration":860,"passed_time":27.57211091,"remaining_time":36.47460432,"test":[0.6703829482]},
-{"learn":[0.6521841478],"iteration":861,"passed_time":27.59555719,"remaining_time":36.43125764,"test":[0.6703818491]},
-{"learn":[0.6521657946],"iteration":862,"passed_time":27.6272049,"remaining_time":36.39876242,"test":[0.6703826129]},
-{"learn":[0.6521304278],"iteration":863,"passed_time":27.65462267,"remaining_time":36.36070759,"test":[0.6703834487]},
-{"learn":[0.6521045712],"iteration":864,"passed_time":27.68321566,"remaining_time":36.3242194,"test":[0.6703868275]},
-{"learn":[0.6520753696],"iteration":865,"passed_time":27.71151671,"remaining_time":36.28736714,"test":[0.6703853357]},
-{"learn":[0.6520519528],"iteration":866,"passed_time":27.73884016,"remaining_time":36.2492571,"test":[0.670450644]},
-{"learn":[0.6520216555],"iteration":867,"passed_time":27.76583897,"remaining_time":36.21074851,"test":[0.6704556991]},
-{"learn":[0.6519926935],"iteration":868,"passed_time":27.79498714,"remaining_time":36.17506382,"test":[0.6704535742]},
-{"learn":[0.6519734186],"iteration":869,"passed_time":27.82082723,"remaining_time":36.13509744,"test":[0.6704495915]}
-]}

ai-engine/catboost_info/learn_error.tsv

@@ -1,871 +0,0 @@
-iter	Logloss
-0	0.692389481
-1	0.6916338586
-2	0.6910159214
-3	0.6903417151
-4	0.6896961461
-5	0.6890979366
-6	0.6884946167
-7	0.6879503686
-8	0.6874528094
-9	0.6869036785
-10	0.6863761921
-11	0.6859038678
-12	0.685410175
-13	0.6849483392
-14	0.6845417792
-15	0.6841038875
-16	0.6836957422
-17	0.6832947461
-18	0.6829014105
-19	0.6825264546
-20	0.6822106577
-21	0.6818649349
-22	0.6815467855
-23	0.6812293319
-24	0.6808837443
-25	0.6805816494
-26	0.6803209634
-27	0.6800350862
-28	0.6797703947
-29	0.6794926675
-30	0.6792251865
-31	0.6789670166
-32	0.678722402
-33	0.678476935
-34	0.6782297335
-35	0.6780226701
-36	0.6778291026
-37	0.6776045324
-38	0.6773969079
-39	0.6771819602
-40	0.6769816736
-41	0.6767984027
-42	0.6766201184
-43	0.6764394377
-44	0.6762698797
-45	0.6760974263
-46	0.6759245179
-47	0.6757673909
-48	0.6756172628
-49	0.675474531
-50	0.6753286933
-51	0.6751900513
-52	0.6750574835
-53	0.6749329567
-54	0.6748033265
-55	0.6746797823
-56	0.674535525
-57	0.6744256514
-58	0.674310819
-59	0.6741967947
-60	0.6740879654
-61	0.6739772476
-62	0.67388281
-63	0.6737789726
-64	0.6736812332
-65	0.6735930009
-66	0.6734947116
-67	0.6733961481
-68	0.6732990195
-69	0.6732133575
-70	0.673111539
-71	0.6730080451
-72	0.6729157861
-73	0.6728347949
-74	0.6727640693
-75	0.6726808811
-76	0.6726029645
-77	0.6725356026
-78	0.6724606887
-79	0.6723849561
-80	0.6723050519
-81	0.6722508802
-82	0.6721773904
-83	0.6721007598
-84	0.6720353564
-85	0.6719790902
-86	0.6719140024
-87	0.6718573633
-88	0.671795602
-89	0.6717369134
-90	0.6716711079
-91	0.6716070843
-92	0.6715517232
-93	0.6714957378
-94	0.6714364567
-95	0.6713881758
-96	0.6713336502
-97	0.6712700267
-98	0.6712154424
-99	0.6711600413
-100	0.6711060533
-101	0.6710494943
-102	0.6709936897
-103	0.6709472183
-104	0.6708914508
-105	0.6708388195
-106	0.6707885854
-107	0.6707454167
-108	0.6706973013
-109	0.6706577031
-110	0.67061108
-111	0.6705625485
-112	0.6705146484
-113	0.6704704423
-114	0.6704155922
-115	0.6703687117
-116	0.6703324232
-117	0.6702884624
-118	0.670253478
-119	0.6702140804
-120	0.6701682529
-121	0.6701320588
-122	0.6700939824
-123	0.6700655902
-124	0.6700190743
-125	0.6699792296
-126	0.6699379404
-127	0.669895454
-128	0.6698563938
-129	0.6698215571
-130	0.6697857067
-131	0.6697449303
-132	0.6697052425
-133	0.6696695553
-134	0.6696269265
-135	0.6695969271
-136	0.6695489786
-137	0.6695173859
-138	0.6694811164
-139	0.6694477439
-140	0.6694082161
-141	0.6693679185
-142	0.6693341916
-143	0.6692933159
-144	0.6692619696
-145	0.6692229289
-146	0.6691840164
-147	0.6691581406
-148	0.6691177196
-149	0.6690851126
-150	0.6690518144
-151	0.6690149711
-152	0.668993877
-153	0.6689596579
-154	0.6689372651
-155	0.6689003045
-156	0.6688680182
-157	0.6688348164
-158	0.6687947046
-159	0.6687605251
-160	0.668726253
-161	0.6686862718
-162	0.668663478
-163	0.6686399521
-164	0.6686058279
-165	0.6685761282
-166	0.6685469327
-167	0.6685157003
-168	0.6684805143
-169	0.6684485765
-170	0.6684144429
-171	0.6683849752
-172	0.6683568537
-173	0.6683266628
-174	0.6682937842
-175	0.6682657097
-176	0.6682301443
-177	0.6681995916
-178	0.6681658267
-179	0.6681422687
-180	0.6681216601
-181	0.6680899019
-182	0.6680676394
-183	0.6680413672
-184	0.6680088406
-185	0.6679873982
-186	0.6679663544
-187	0.6679417375
-188	0.6679100197
-189	0.667881208
-190	0.6678475427
-191	0.6678310341
-192	0.6678060257
-193	0.6677789336
-194	0.6677478773
-195	0.6677212408
-196	0.667704316
-197	0.6676819639
-198	0.6676554448
-199	0.6676318346
-200	0.6676074705
-201	0.6675849784
-202	0.6675631744
-203	0.6675397619
-204	0.6675169086
-205	0.6674864762
-206	0.6674670714
-207	0.6674375599
-208	0.6674148457
-209	0.6673974446
-210	0.6673812139
-211	0.6673515687
-212	0.6673197956
-213	0.6672900754
-214	0.6672550009
-215	0.6672271563
-216	0.667204521
-217	0.667181968
-218	0.6671640023
-219	0.66714351
-220	0.6671167156
-221	0.6670915937
-222	0.6670595279
-223	0.667033994
-224	0.6670008246
-225	0.6669858319
-226	0.6669553964
-227	0.6669274683
-228	0.666896348
-229	0.6668698686
-230	0.6668513411
-231	0.6668309985
-232	0.6668058585
-233	0.6667845908
-234	0.6667582863
-235	0.6667332943
-236	0.6667070085
-237	0.6666907315
-238	0.6666633028
-239	0.6666406707
-240	0.6666134624
-241	0.6665850522
-242	0.6665631193
-243	0.6665412643
-244	0.6665168385
-245	0.6664904845
-246	0.6664678274
-247	0.6664539777
-248	0.6664334121
-249	0.6664121724
-250	0.666392034
-251	0.666366899
-252	0.6663414098
-253	0.6663157816
-254	0.6662989799
-255	0.6662696102
-256	0.6662479711
-257	0.6662231874
-258	0.6661947927
-259	0.6661669951
-260	0.6661426137
-261	0.6661216749
-262	0.6660983123
-263	0.6660803402
-264	0.6660617842
-265	0.6660443878
-266	0.6660176079
-267	0.6659967546
-268	0.6659751467
-269	0.6659539329
-270	0.6659263951
-271	0.6659038921
-272	0.6658767418
-273	0.6658510507
-274	0.6658210119
-275	0.6657963011
-276	0.6657748552
-277	0.6657490013
-278	0.665732402
-279	0.6657118786
-280	0.665684467
-281	0.6656584634
-282	0.6656309991
-283	0.6656073482
-284	0.6655890957
-285	0.6655665563
-286	0.6655452454
-287	0.6655255286
-288	0.6655053548
-289	0.6654893396
-290	0.6654648912
-291	0.6654442759
-292	0.6654173127
-293	0.6653914518
-294	0.6653648946
-295	0.665344141
-296	0.6653140817
-297	0.665295365
-298	0.6652787488
-299	0.6652502991
-300	0.665231168
-301	0.6652136682
-302	0.6651903001
-303	0.6651697153
-304	0.6651525958
-305	0.6651322685
-306	0.6651113828
-307	0.6650886807
-308	0.6650622251
-309	0.6650429987
-310	0.665015513
-311	0.6650019022
-312	0.664979951
-313	0.6649549638
-314	0.6649340455
-315	0.6649162445
-316	0.6649048119
-317	0.6648796463
-318	0.6648605481
-319	0.6648429084
-320	0.6648238121
-321	0.6647969527
-322	0.6647854723
-323	0.6647589304
-324	0.6647429024
-325	0.6647237508
-326	0.6647059396
-327	0.664686288
-328	0.6646532527
-329	0.6646306438
-330	0.6646098516
-331	0.6645858284
-332	0.6645707188
-333	0.6645485788
-334	0.6645305696
-335	0.6645108881
-336	0.6644923286
-337	0.6644805222
-338	0.6644572776
-339	0.6644320741
-340	0.6644115048
-341	0.6643949013
-342	0.6643619789
-343	0.6643389502
-344	0.6643088915
-345	0.664286972
-346	0.664274149
-347	0.6642536926
-348	0.6642357634
-349	0.664207914
-350	0.6641853097
-351	0.6641654917
-352	0.664143804
-353	0.6641290647
-354	0.6641117244
-355	0.6640880219
-356	0.6640669415
-357	0.6640462999
-358	0.664030296
-359	0.6640028542
-360	0.6639813347
-361	0.6639597941
-362	0.6639429832
-363	0.6639222708
-364	0.6639065546
-365	0.6638823236
-366	0.6638648195
-367	0.6638436235
-368	0.6638208732
-369	0.6637956357
-370	0.6637718453
-371	0.663756918
-372	0.6637353525
-373	0.6637143112
-374	0.6636956547
-375	0.663680995
-376	0.66366728
-377	0.6636487567
-378	0.6636266904
-379	0.6636116064
-380	0.6635902746
-381	0.6635654896
-382	0.6635393029
-383	0.6635171734
-384	0.663500789
-385	0.663477743
-386	0.6634584806
-387	0.6634337499
-388	0.6634135584
-389	0.6633868455
-390	0.6633755323
-391	0.663356103
-392	0.6633337631
-393	0.663319422
-394	0.6632911566
-395	0.6632687875
-396	0.6632431997
-397	0.6632189331
-398	0.663201035
-399	0.6631898553
-400	0.6631712482
-401	0.663143025
-402	0.663121538
-403	0.6631087792
-404	0.6630859067
-405	0.663066483
-406	0.6630443652
-407	0.6630250376
-408	0.6630007822
-409	0.6629768728
-410	0.6629528093
-411	0.6629260936
-412	0.6629102182
-413	0.6628863488
-414	0.6628648972
-415	0.6628454339
-416	0.6628200274
-417	0.6627942591
-418	0.6627744647
-419	0.662765485
-420	0.6627503257
-421	0.6627323029
-422	0.6627111509
-423	0.6626785863
-424	0.6626576561
-425	0.6626363113
-426	0.6626181065
-427	0.66259794
-428	0.6625765658
-429	0.6625526572
-430	0.66253135
-431	0.6625035695
-432	0.662480212
-433	0.6624611632
-434	0.6624332625
-435	0.6624120584
-436	0.6623941719
-437	0.6623766304
-438	0.6623623329
-439	0.6623442925
-440	0.6623212715
-441	0.6623025941
-442	0.6622749791
-443	0.6622534499
-444	0.6622305473
-445	0.6622059333
-446	0.6621871707
-447	0.6621638454
-448	0.6621511296
-449	0.6621349978
-450	0.6621120424
-451	0.6620958271
-452	0.6620793528
-453	0.6620572713
-454	0.6620395025
-455	0.6620188044
-456	0.6620017347
-457	0.6619811454
-458	0.6619695569
-459	0.661952377
-460	0.6619237442
-461	0.6619089407
-462	0.6618886168
-463	0.6618831383
-464	0.6618690774
-465	0.661845878
-466	0.6618290213
-467	0.6618050064
-468	0.6617832833
-469	0.6617652311
-470	0.6617443144
-471	0.6617202619
-472	0.6617005831
-473	0.6616824419
-474	0.6616538226
-475	0.6616314155
-476	0.6616127861
-477	0.6616029072
-478	0.6615843751
-479	0.661563216
-480	0.6615432257
-481	0.6615263324
-482	0.6615033259
-483	0.661484293
-484	0.6614678231
-485	0.6614463024
-486	0.6614155436
-487	0.6613958945
-488	0.661380611
-489	0.6613677802
-490	0.6613530086
-491	0.6613248211
-492	0.6613059359
-493	0.6612729965
-494	0.6612624948
-495	0.6612401679
-496	0.6612191637
-497	0.6611912219
-498	0.6611773017
-499	0.6611638216
-500	0.6611450533
-501	0.6611179111
-502	0.6610959069
-503	0.6610728788
-504	0.6610436668
-505	0.6610188976
-506	0.6610030555
-507	0.6609831174
-508	0.6609586562
-509	0.660935882
-510	0.6609202024
-511	0.6609011137
-512	0.6608726737
-513	0.6608608849
-514	0.6608387256
-515	0.6608136063
-516	0.6607946343
-517	0.6607703935
-518	0.6607509625
-519	0.6607238109
-520	0.6606999858
-521	0.6606813873
-522	0.6606610372
-523	0.660638456
-524	0.6606156483
-525	0.6605968623
-526	0.6605735776
-527	0.6605517294
-528	0.6605309239
-529	0.6605086434
-530	0.6604803349
-531	0.6604566326
-532	0.6604430839
-533	0.6604273738
-534	0.6604048016
-535	0.6603845173
-536	0.6603669212
-537	0.6603488983
-538	0.6603176881
-539	0.6602953862
-540	0.6602672025
-541	0.6602568636
-542	0.660235705
-543	0.6602152295
-544	0.6601897709
-545	0.6601683731
-546	0.6601472267
-547	0.6601262337
-548	0.6601119991
-549	0.6600869973
-550	0.6600667497
-551	0.6600397508
-552	0.660016863
-553	0.6599933158
-554	0.6599632649
-555	0.6599446007
-556	0.6599138126
-557	0.6598965504
-558	0.6598785723
-559	0.659860838
-560	0.6598408724
-561	0.6598244857
-562	0.6598082469
-563	0.6597851673
-564	0.6597683521
-565	0.6597479006
-566	0.6597310938
-567	0.6597096581
-568	0.6596862311
-569	0.6596574779
-570	0.6596385418
-571	0.6596189903
-572	0.65959275
-573	0.6595730662
-574	0.6595566809
-575	0.6595365076
-576	0.6595163446
-577	0.6594816637
-578	0.6594570142
-579	0.6594353055
-580	0.6594162362
-581	0.659395036
-582	0.6593798831
-583	0.6593556719
-584	0.6593292627
-585	0.6592976737
-586	0.6592754841
-587	0.6592510441
-588	0.6592290326
-589	0.6592097404
-590	0.6591876204
-591	0.6591705995
-592	0.6591456195
-593	0.6591107122
-594	0.6590819533
-595	0.6590551327
-596	0.6590373916
-597	0.6590177149
-598	0.6589946095
-599	0.6589697628
-600	0.6589442269
-601	0.6589182437
-602	0.6588837179
-603	0.6588674101
-604	0.6588406916
-605	0.6588149945
-606	0.6587866031
-607	0.6587636648
-608	0.6587502469
-609	0.6587292784
-610	0.6587104112
-611	0.6586953782
-612	0.6586641191
-613	0.6586450136
-614	0.6586136263
-615	0.6585862768
-616	0.6585585235
-617	0.6585371631
-618	0.6585092632
-619	0.6584914317
-620	0.6584662432
-621	0.6584454668
-622	0.6584249408
-623	0.6583931228
-624	0.6583660767
-625	0.658354264
-626	0.6583253625
-627	0.6582968632
-628	0.6582687399
-629	0.658242535
-630	0.6582199874
-631	0.6581918101
-632	0.6581735218
-633	0.6581445869
-634	0.6581202427
-635	0.6580977862
-636	0.6580724179
-637	0.6580426322
-638	0.6580111256
-639	0.6579834747
-640	0.6579541367
-641	0.6579254503
-642	0.657898555
-643	0.6578676875
-644	0.6578324163
-645	0.6578062223
-646	0.6577760631
-647	0.6577483474
-648	0.6577249642
-649	0.6576974966
-650	0.657675114
-651	0.6576447891
-652	0.6576102356
-653	0.6575793887
-654	0.6575543309
-655	0.6575340787
-656	0.6575061464
-657	0.657476113
-658	0.6574447014
-659	0.6574247361
-660	0.6574034983
-661	0.6573783832
-662	0.657357694
-663	0.6573411592
-664	0.6573118559
-665	0.6572819076
-666	0.6572430097
-667	0.6572160391
-668	0.6571931413
-669	0.6571737099
-670	0.6571532872
-671	0.6571208939
-672	0.6570887673
-673	0.6570633692
-674	0.6570454361
-675	0.6570231031
-676	0.6570052089
-677	0.6569855794
-678	0.6569579709
-679	0.6569333354
-680	0.6569069617
-681	0.6568931857
-682	0.6568734532
-683	0.6568435196
-684	0.6568108038
-685	0.6567811374
-686	0.6567467284
-687	0.6567172734
-688	0.6566967606
-689	0.6566720128
-690	0.6566441608
-691	0.6566172287
-692	0.6565952549
-693	0.6565702687
-694	0.6565392213
-695	0.6565157938
-696	0.6564902789
-697	0.6564644734
-698	0.6564349549
-699	0.6564046572
-700	0.6563744107
-701	0.6563525063
-702	0.6563189867
-703	0.6562939062
-704	0.6562739297
-705	0.656256438
-706	0.6562366475
-707	0.6562073096
-708	0.6561864222
-709	0.6561578826
-710	0.6561208567
-711	0.6560924703
-712	0.6560656907
-713	0.6560362588
-714	0.6560124527
-715	0.6559875055
-716	0.6559547281
-717	0.6559230866
-718	0.6558924823
-719	0.6558676469
-720	0.6558459277
-721	0.6558149638
-722	0.6557812248
-723	0.6557546502
-724	0.6557274948
-725	0.6557044723
-726	0.6556751811
-727	0.6556539158
-728	0.6556182915
-729	0.6555977079
-730	0.6555667903
-731	0.6555394075
-732	0.6555122742
-733	0.6554814941
-734	0.6554517373
-735	0.655429552
-736	0.655396579
-737	0.6553735864
-738	0.6553472597
-739	0.6553252832
-740	0.6552971659
-741	0.6552763852
-742	0.6552488203
-743	0.65521229
-744	0.6551949744
-745	0.6551673797
-746	0.6551421856
-747	0.6551255516
-748	0.6551019608
-749	0.6550758728
-750	0.655051966
-751	0.6550351058
-752	0.6549998756
-753	0.6549721212
-754	0.6549401744
-755	0.6549207325
-756	0.6548900891
-757	0.6548682731
-758	0.6548418938
-759	0.6548234717
-760	0.6547996833
-761	0.6547726174
-762	0.6547509314
-763	0.6547168175
-764	0.6546907846
-765	0.6546671611
-766	0.6546475893
-767	0.6546206223
-768	0.6545874193
-769	0.6545620629
-770	0.6545346297
-771	0.6545172316
-772	0.6544943049
-773	0.6544632323
-774	0.6544384097
-775	0.6544084745
-776	0.6543765257
-777	0.6543536123
-778	0.6543303593
-779	0.6543005831
-780	0.6542678123
-781	0.6542439303
-782	0.6542100401
-783	0.6541836178
-784	0.654158129
-785	0.6541343464
-786	0.6541092921
-787	0.6540812254
-788	0.654060259
-789	0.6540467253
-790	0.6540306837
-791	0.6540103667
-792	0.6539821302
-793	0.6539577914
-794	0.653923724
-795	0.6539086888
-796	0.6538798424
-797	0.6538566996
-798	0.6538290752
-799	0.6538051255
-800	0.6537917354
-801	0.6537684302
-802	0.6537402991
-803	0.6537165427
-804	0.6536853601
-805	0.6536681479
-806	0.6536409101
-807	0.6536120189
-808	0.6535912493
-809	0.6535617421
-810	0.6535315174
-811	0.6534972927
-812	0.6534818476
-813	0.6534498323
-814	0.6534305025
-815	0.6534081059
-816	0.6533765804
-817	0.6533441549
-818	0.6533053405
-819	0.6532838469
-820	0.6532604302
-821	0.6532364412
-822	0.6532100089
-823	0.6531782515
-824	0.6531449701
-825	0.653115452
-826	0.6530787602
-827	0.653052397
-828	0.6530313579
-829	0.6530010363
-830	0.6529752146
-831	0.652954801
-832	0.6529330351
-833	0.6528993709
-834	0.6528665883
-835	0.6528413041
-836	0.6528217161
-837	0.6527978782
-838	0.6527789461
-839	0.6527432001
-840	0.6527139767
-841	0.6526857244
-842	0.652657086
-843	0.6526355016
-844	0.6526054936
-845	0.6525793707
-846	0.6525584692
-847	0.6525279747
-848	0.6525038765
-849	0.6524849104
-850	0.6524610603
-851	0.6524357337
-852	0.6524082286
-853	0.65238051
-854	0.6523557826
-855	0.6523391233
-856	0.652325347
-857	0.6522924958
-858	0.6522623584
-859	0.6522343891
-860	0.6522094424
-861	0.6521841478
-862	0.6521657946
-863	0.6521304278
-864	0.6521045712
-865	0.6520753696
-866	0.6520519528
-867	0.6520216555
-868	0.6519926935
-869	0.6519734186

ai-engine/catboost_info/time_left.tsv

@@ -1,871 +0,0 @@
-iter	Passed	Remaining
-0	46	93548
-1	83	83419
-2	132	88415
-3	162	81250
-4	196	78573
-5	230	76747
-6	269	76701
-7	319	79674
-8	364	80653
-9	411	81918
-10	456	82497
-11	491	81432
-12	522	79809
-13	555	78774
-14	595	78777
-15	630	78123
-16	662	77290
-17	700	77124
-18	730	76120
-19	764	75651
-20	804	75774
-21	835	75128
-22	886	76169
-23	920	75764
-24	960	75853
-25	989	75130
-26	1025	74941
-27	1060	74714
-28	1104	75079
-29	1141	74976
-30	1180	74975
-31	1213	74640
-32	1245	74260
-33	1287	74434
-34	1327	74528
-35	1376	75071
-36	1427	75741
-37	1468	75804
-38	1508	75857
-39	1549	75922
-40	1586	75781
-41	1621	75590
-42	1663	75705
-43	1701	75621
-44	1739	75591
-45	1776	75460
-46	1819	75616
-47	1869	76025
-48	1916	76288
-49	1953	76191
-50	1993	76197
-51	2038	76381
-52	2080	76420
-53	2158	77788
-54	2220	78529
-55	2286	79390
-56	2328	79372
-57	2367	79254
-58	2409	79257
-59	2444	79049
-60	2484	78985
-61	2521	78820
-62	2554	78528
-63	2593	78466
-64	2623	78111
-65	2660	77969
-66	2695	77776
-67	2725	77446
-68	2761	77291
-69	2791	76975
-70	2824	76739
-71	2861	76611
-72	2897	76476
-73	2935	76408
-74	3040	78027
-75	3097	78411
-76	3152	78741
-77	3216	79248
-78	3256	79195
-79	3305	79336
-80	3348	79320
-81	3381	79089
-82	3416	78911
-83	3480	79399
-84	3535	79649
-85	3581	79716
-86	3612	79428
-87	3644	79185
-88	3678	78975
-89	3712	78785
-90	3743	78531
-91	3775	78297
-92	3806	78047
-93	3837	77821
-94	3871	77629
-95	3913	77618
-96	3945	77403
-97	3989	77433
-98	4020	77204
-99	4053	77020
-100	4084	76789
-101	4116	76597
-102	4148	76401
-103	4176	76141
-104	4202	75845
-105	4232	75634
-106	4261	75390
-107	4290	75168
-108	4324	75018
-109	4351	74766
-110	4386	74648
-111	4424	74577
-112	4458	74455
-113	4497	74400
-114	4533	74307
-115	4564	74136
-116	4596	73981
-117	4628	73818
-118	4668	73786
-119	4692	73509
-120	4723	73354
-121	4756	73220
-122	4788	73065
-123	4815	72854
-124	4843	72647
-125	4875	72514
-126	4916	72515
-127	4952	72436
-128	4991	72397
-129	5028	72327
-130	5059	72180
-131	5096	72116
-132	5125	71946
-133	5156	71804
-134	5190	71704
-135	5221	71564
-136	5251	71407
-137	5274	71165
-138	5309	71084
-139	5344	71008
-140	5377	70902
-141	5416	70866
-142	5452	70803
-143	5490	70760
-144	5521	70641
-145	5553	70522
-146	5582	70365
-147	5611	70217
-148	5636	70026
-149	5673	69975
-150	5706	69874
-151	5738	69764
-152	5765	69605
-153	5795	69471
-154	5817	69246
-155	5853	69191
-156	5888	69122
-157	5924	69070
-158	5964	69061
-159	5996	68963
-160	6022	68789
-161	6050	68650
-162	6079	68510
-163	6108	68385
-164	6140	68292
-165	6169	68162
-166	6202	68074
-167	6231	67953
-168	6263	67858
-169	6295	67764
-170	6325	67656
-171	6356	67561
-172	6395	67545
-173	6437	67554
-174	6472	67495
-175	6503	67395
-176	6533	67291
-177	6562	67174
-178	6590	67049
-179	6624	66982
-180	6655	66882
-181	6687	66804
-182	6718	66703
-183	6751	66632
-184	6784	66559
-185	6810	66424
-186	6832	66246
-187	6867	66187
-188	6918	66294
-189	6969	66393
-190	7018	66470
-191	7074	66614
-192	7117	66635
-193	7191	66943
-194	7242	67036
-195	7282	67027
-196	7317	66967
-197	7351	66903
-198	7389	66879
-199	7432	66896
-200	7471	66869
-201	7506	66814
-202	7540	66752
-203	7568	66628
-204	7605	66596
-205	7638	66519
-206	7665	66397
-207	7700	66340
-208	7734	66276
-209	7766	66197
-210	7796	66106
-211	7831	66053
-212	7871	66037
-213	7910	66016
-214	7951	66014
-215	7989	65983
-216	8025	65946
-217	8058	65872
-218	8087	65768
-219	8112	65638
-220	8148	65594
-221	8197	65655
-222	8239	65655
-223	8268	65556
-224	8298	65466
-225	8327	65366
-226	8357	65278
-227	8384	65167
-228	8418	65103
-229	8453	65058
-230	8490	65020
-231	8523	64958
-232	8550	64848
-233	8575	64718
-234	8607	64648
-235	8635	64545
-236	8660	64426
-237	8691	64345
-238	8719	64250
-239	8746	64137
-240	8773	64038
-241	8803	63951
-242	8833	63873
-243	8862	63779
-244	8892	63698
-245	8932	63688
-246	8962	63611
-247	8991	63521
-248	9021	63442
-249	9051	63358
-250	9085	63306
-251	9110	63193
-252	9137	63093
-253	9174	63066
-254	9196	62935
-255	9238	62934
-256	9267	62855
-257	9297	62776
-258	9324	62681
-259	9357	62625
-260	9388	62552
-261	9427	62536
-262	9461	62491
-263	9496	62443
-264	9524	62356
-265	9553	62278
-266	9590	62247
-267	9620	62172
-268	9645	62071
-269	9682	62040
-270	9711	61962
-271	9739	61872
-272	9768	61797
-273	9804	61761
-274	9848	61777
-275	9886	61755
-276	9925	61740
-277	9965	61728
-278	9995	61656
-279	10022	61564
-280	10055	61516
-281	10080	61410
-282	10111	61344
-283	10147	61311
-284	10175	61230
-285	10202	61141
-286	10234	61084
-287	10264	61018
-288	10299	60977
-289	10323	60874
-290	10353	60804
-291	10394	60803
-292	10431	60773
-293	10471	60763
-294	10503	60707
-295	10534	60645
-296	10576	60646
-297	10612	60612
-298	10639	60525
-299	10668	60453
-300	10702	60411
-301	10729	60326
-302	10764	60290
-303	10801	60263
-304	10829	60182
-305	10857	60108
-306	10892	60067
-307	10930	60047
-308	10972	60045
-309	11002	59983
-310	11030	59902
-311	11058	59828
-312	11092	59788
-313	11117	59696
-314	11149	59641
-315	11187	59617
-316	11211	59525
-317	11243	59468
-318	11274	59413
-319	11304	59346
-320	11334	59287
-321	11362	59209
-322	11394	59158
-323	11436	59158
-324	11477	59153
-325	11513	59122
-326	11547	59081
-327	11572	58991
-328	11607	58956
-329	11637	58894
-330	11668	58833
-331	11700	58785
-332	11724	58694
-333	11757	58648
-334	11780	58550
-335	11815	58515
-336	11844	58451
-337	11869	58364
-338	11905	58335
-339	11941	58302
-340	11986	58315
-341	12020	58274
-342	12066	58292
-343	12122	58358
-344	12177	58415
-345	12221	58422
-346	12264	58423
-347	12300	58394
-348	12324	58304
-349	12354	58243
-350	12401	58262
-351	12438	58232
-352	12479	58228
-353	12512	58179
-354	12541	58116
-355	12569	58044
-356	12597	57977
-357	12628	57920
-358	12653	57839
-359	12682	57775
-360	12720	57752
-361	12744	57666
-362	12770	57592
-363	12811	57583
-364	12841	57522
-365	12870	57460
-366	12897	57386
-367	12938	57378
-368	12974	57347
-369	13009	57313
-370	13038	57249
-371	13078	57235
-372	13117	57216
-373	13147	57159
-374	13181	57118
-375	13205	57036
-376	13235	56979
-377	13274	56960
-378	13306	56911
-379	13333	56841
-380	13366	56798
-381	13396	56741
-382	13421	56666
-383	13467	56674
-384	13508	56664
-385	13540	56616
-386	13569	56559
-387	13598	56496
-388	13627	56438
-389	13656	56376
-390	13685	56317
-391	13717	56271
-392	13750	56227
-393	13771	56135
-394	13804	56090
-395	13825	55999
-396	13858	55957
-397	13888	55904
-398	13917	55843
-399	13953	55812
-400	13994	55802
-401	14025	55752
-402	14048	55670
-403	14076	55607
-404	14105	55551
-405	14142	55526
-406	14182	55511
-407	14214	55464
-408	14240	55394
-409	14267	55328
-410	14299	55284
-411	14324	55213
-412	14351	55146
-413	14379	55086
-414	14410	55036
-415	14451	55025
-416	14484	54984
-417	14513	54929
-418	14536	54851
-419	14565	54793
-420	14587	54710
-421	14615	54650
-422	14642	54588
-423	14666	54515
-424	14690	54441
-425	14719	54384
-426	14739	54297
-427	14772	54257
-428	14790	54164
-429	14824	54125
-430	14844	54039
-431	14876	53995
-432	14906	53946
-433	14938	53902
-434	14980	53894
-435	15006	53829
-436	15033	53770
-437	15059	53706
-438	15085	53639
-439	15110	53574
-440	15134	53503
-441	15160	53438
-442	15184	53369
-443	15211	53308
-444	15234	53236
-445	15266	53193
-446	15287	53114
-447	15316	53059
-448	15336	52978
-449	15366	52929
-450	15393	52870
-451	15429	52843
-452	15469	52828
-453	15490	52748
-454	15523	52712
-455	15550	52653
-456	15577	52594
-457	15604	52536
-458	15630	52476
-459	15656	52414
-460	15682	52353
-461	15711	52304
-462	15736	52238
-463	15765	52188
-464	15786	52112
-465	15817	52068
-466	15839	51996
-467	15873	51961
-468	15903	51916
-469	15935	51873
-470	15969	51840
-471	15994	51779
-472	16022	51726
-473	16047	51663
-474	16073	51605
-475	16099	51546
-476	16128	51495
-477	16152	51431
-478	16176	51367
-479	16205	51317
-480	16228	51250
-481	16255	51194
-482	16277	51123
-483	16305	51071
-484	16328	51005
-485	16362	50973
-486	16392	50928
-487	16426	50894
-488	16459	50860
-489	16480	50787
-490	16510	50743
-491	16530	50668
-492	16561	50625
-493	16585	50562
-494	16613	50510
-495	16638	50453
-496	16663	50393
-497	16690	50339
-498	16716	50282
-499	16740	50222
-500	16773	50186
-501	16802	50139
-502	16836	50107
-503	16873	50085
-504	16921	50094
-505	16989	50163
-506	17038	50173
-507	17069	50132
-508	17110	50121
-509	17145	50091
-510	17190	50091
-511	17219	50044
-512	17247	49994
-513	17271	49932
-514	17298	49878
-515	17343	49878
-516	17373	49836
-517	17417	49831
-518	17460	49823
-519	17490	49781
-520	17518	49731
-521	17546	49680
-522	17571	49622
-523	17600	49577
-524	17625	49520
-525	17655	49474
-526	17679	49414
-527	17707	49366
-528	17729	49300
-529	17758	49254
-530	17781	49191
-531	17808	49141
-532	17829	49071
-533	17862	49038
-534	17905	49031
-535	18028	49241
-536	18072	49236
-537	18106	49203
-538	18135	49157
-539	18165	49114
-540	18200	49083
-541	18223	49022
-542	18254	48980
-543	18280	48927
-544	18307	48876
-545	18338	48834
-546	18367	48790
-547	18411	48783
-548	18444	48747
-549	18470	48693
-550	18503	48660
-551	18531	48611
-552	18557	48558
-553	18584	48508
-554	18625	48493
-555	18650	48436
-556	18677	48388
-557	18703	48333
-558	18729	48282
-559	18756	48231
-560	18781	48176
-561	18808	48126
-562	18834	48074
-563	18869	48043
-564	18902	48008
-565	18930	47960
-566	18958	47914
-567	18983	47859
-568	19016	47824
-569	19037	47761
-570	19068	47720
-571	19090	47660
-572	19111	47595
-573	19141	47553
-574	19164	47494
-575	19196	47458
-576	19217	47393
-577	19249	47358
-578	19274	47303
-579	19298	47247
-580	19324	47195
-581	19357	47162
-582	19391	47130
-583	19427	47103
-584	19460	47070
-585	19483	47012
-586	19511	46967
-587	19542	46929
-588	19564	46867
-589	19597	46833
-590	19621	46779
-591	19647	46729
-592	19670	46672
-593	19699	46627
-594	19726	46582
-595	19753	46532
-596	19778	46480
-597	19803	46429
-598	19830	46381
-599	19857	46335
-600	19896	46313
-601	19925	46271
-602	19957	46236
-603	19991	46204
-604	20019	46159
-605	20047	46115
-606	20072	46063
-607	20098	46015
-608	20123	45963
-609	20149	45913
-610	20176	45867
-611	20202	45817
-612	20230	45774
-613	20253	45719
-614	20285	45682
-615	20307	45626
-616	20338	45589
-617	20361	45532
-618	20394	45500
-619	20423	45459
-620	20454	45420
-621	20488	45390
-622	20510	45333
-623	20543	45301
-624	20569	45252
-625	20594	45201
-626	20619	45151
-627	20646	45107
-628	20675	45066
-629	20701	45016
-630	20727	44970
-631	20752	44919
-632	20782	44881
-633	20804	44825
-634	20837	44791
-635	20862	44742
-636	20892	44704
-637	20931	44683
-638	20960	44643
-639	20994	44612
-640	21022	44570
-641	21052	44531
-642	21082	44493
-643	21107	44443
-644	21135	44401
-645	21160	44351
-646	21185	44302
-647	21210	44253
-648	21236	44208
-649	21262	44161
-650	21288	44113
-651	21315	44068
-652	21343	44027
-653	21377	43997
-654	21403	43949
-655	21440	43926
-656	21477	43903
-657	21502	43854
-658	21533	43819
-659	21559	43772
-660	21586	43727
-661	21611	43680
-662	21637	43633
-663	21662	43586
-664	21688	43539
-665	21714	43493
-666	21742	43451
-667	21771	43413
-668	21818	43409
-669	21846	43366
-670	21888	43352
-671	21934	43345
-672	21971	43322
-673	22019	43320
-674	22053	43289
-675	22090	43266
-676	22141	43269
-677	22176	43240
-678	22213	43215
-679	22239	43171
-680	22270	43134
-681	22296	43088
-682	22321	43041
-683	22350	43002
-684	22379	42962
-685	22419	42944
-686	22452	42912
-687	22484	42878
-688	22511	42834
-689	22537	42789
-690	22571	42757
-691	22598	42714
-692	22624	42669
-693	22653	42630
-694	22680	42586
-695	22708	42545
-696	22739	42510
-697	22761	42457
-698	22792	42421
-699	22816	42373
-700	22845	42333
-701	22870	42288
-702	22902	42253
-703	22942	42234
-704	22974	42201
-705	23002	42160
-706	23033	42124
-707	23054	42071
-708	23086	42038
-709	23115	41999
-710	23143	41957
-711	23169	41914
-712	23195	41868
-713	23230	41840
-714	23259	41801
-715	23287	41760
-716	23311	41713
-717	23341	41676
-718	23372	41641
-719	23405	41610
-720	23438	41578
-721	23483	41566
-722	23507	41519
-723	23540	41488
-724	23566	41444
-725	23595	41406
-726	23623	41365
-727	23648	41320
-728	23677	41281
-729	23700	41231
-730	23728	41192
-731	23752	41144
-732	23784	41111
-733	23807	41063
-734	23840	41031
-735	23870	40994
-736	23908	40972
-737	23941	40940
-738	23974	40909
-739	24006	40875
-740	24036	40838
-741	24064	40798
-742	24092	40759
-743	24127	40730
-744	24153	40688
-745	24179	40644
-746	24207	40604
-747	24233	40561
-748	24261	40522
-749	24295	40491
-750	24318	40444
-751	24349	40410
-752	24376	40368
-753	24408	40335
-754	24442	40306
-755	24474	40273
-756	24508	40242
-757	24548	40222
-758	24575	40182
-759	24605	40145
-760	24632	40104
-761	24660	40064
-762	24689	40027
-763	24714	39982
-764	24745	39949
-765	24766	39897
-766	24797	39863
-767	24825	39823
-768	24854	39786
-769	24880	39744
-770	24909	39706
-771	24940	39672
-772	24970	39635
-773	25004	39606
-774	25030	39564
-775	25056	39522
-776	25086	39486
-777	25107	39436
-778	25139	39403
-779	25159	39351
-780	25188	39314
-781	25214	39272
-782	25240	39230
-783	25266	39188
-784	25288	39141
-785	25315	39101
-786	25341	39058
-787	25367	39016
-788	25391	38972
-789	25417	38930
-790	25448	38895
-791	25482	38867
-792	25514	38834
-793	25542	38795
-794	25569	38756
-795	25595	38714
-796	25618	38669
-797	25643	38626
-798	25667	38581
-799	25695	38543
-800	25716	38494
-801	25743	38454
-802	25770	38415
-803	25790	38364
-804	25822	38332
-805	25843	38284
-806	25873	38249
-807	25896	38203
-808	25925	38167
-809	25955	38131
-810	25988	38101
-811	26028	38080
-812	26055	38042
-813	26081	38000
-814	26108	37961
-815	26131	37916
-816	26159	37878
-817	26188	37841
-818	26214	37800
-819	26242	37764
-820	26272	37728
-821	26298	37688
-822	26327	37652
-823	26359	37619
-824	26385	37580
-825	26408	37534
-826	26444	37507
-827	26477	37478
-828	26517	37456
-829	26539	37411
-830	26573	37382
-831	26597	37339
-832	26623	37298
-833	26650	37259
-834	26677	37221
-835	26704	37182
-836	26728	37138
-837	26763	37111
-838	26791	37073
-839	26822	37041
-840	26872	37033
-841	26924	37029
-842	26982	37033
-843	27054	37055
-844	27097	37038
-845	27120	36994
-846	27146	36954
-847	27180	36925
-848	27206	36884
-849	27234	36846
-850	27260	36807
-851	27289	36770
-852	27318	36734
-853	27347	36698
-854	27386	36675
-855	27413	36637
-856	27439	36596
-857	27471	36564
-858	27501	36529
-859	27535	36500
-860	27572	36474
-861	27595	36431
-862	27627	36398
-863	27654	36360
-864	27683	36324
-865	27711	36287
-866	27738	36249
-867	27765	36210
-868	27794	36175
-869	27820	36135

ai-engine/config/market_anchor_corrections.json

@@ -0,0 +1,131 @@
+{
+  "version": "2026-06-10T23:11:49",
+  "fitted_on": {
+    "days": 540,
+    "test_days": 90,
+    "n_train": 64603,
+    "n_test": 14450
+  },
+  "validated": {
+    "ece_home": {
+      "raw": 0.01803,
+      "active_before": 0.01312,
+      "candidate_oos": 0.01301
+    },
+    "ece_away": {
+      "raw": 0.01234,
+      "active_before": 0.01234,
+      "candidate_oos": 0.00845
+    }
+  },
+  "gates": {
+    "min_n": 1500,
+    "shrink": 0.5,
+    "clip": 0.05,
+    "min_delta": 0.004
+  },
+  "corrections": {
+    "ms_home": [
+      {
+        "lo": 0.05,
+        "hi": 0.15,
+        "delta": 0.0,
+        "n": 2124,
+        "raw_gap": -0.0072
+      },
+      {
+        "lo": 0.15,
+        "hi": 0.25,
+        "delta": 0.0,
+        "n": 6476,
+        "raw_gap": -0.0031
+      },
+      {
+        "lo": 0.25,
+        "hi": 0.35,
+        "delta": 0.0,
+        "n": 12565,
+        "raw_gap": -0.0018
+      },
+      {
+        "lo": 0.35,
+        "hi": 0.45,
+        "delta": 0.0,
+        "n": 16431,
+        "raw_gap": 0.006
+      },
+      {
+        "lo": 0.45,
+        "hi": 0.55,
+        "delta": 0.0124,
+        "n": 12995,
+        "raw_gap": 0.0248
+      },
+      {
+        "lo": 0.55,
+        "hi": 0.65,
+        "delta": 0.0154,
+        "n": 8479,
+        "raw_gap": 0.0307
+      },
+      {
+        "lo": 0.65,
+        "hi": 0.75,
+        "delta": 0.0203,
+        "n": 4638,
+        "raw_gap": 0.0407
+      }
+    ],
+    "ms_away": [
+      {
+        "lo": 0.05,
+        "hi": 0.15,
+        "delta": -0.0077,
+        "n": 6762,
+        "raw_gap": -0.0154
+      },
+      {
+        "lo": 0.15,
+        "hi": 0.25,
+        "delta": -0.0048,
+        "n": 16211,
+        "raw_gap": -0.0097
+      },
+      {
+        "lo": 0.25,
+        "hi": 0.35,
+        "delta": 0.0,
+        "n": 18440,
+        "raw_gap": -0.002
+      },
+      {
+        "lo": 0.35,
+        "hi": 0.45,
+        "delta": 0.009,
+        "n": 12061,
+        "raw_gap": 0.0181
+      },
+      {
+        "lo": 0.45,
+        "hi": 0.55,
+        "delta": 0.0116,
+        "n": 5930,
+        "raw_gap": 0.0231
+      },
+      {
+        "lo": 0.55,
+        "hi": 0.65,
+        "delta": 0.0199,
+        "n": 3287,
+        "raw_gap": 0.0399
+      },
+      {
+        "lo": 0.65,
+        "hi": 0.75,
+        "delta": 0.0295,
+        "n": 1580,
+        "raw_gap": 0.0589
+      }
+    ]
+  }
+}

ai-engine/data/league_confidence.json

@@ -0,0 +1,134 @@
+{
+ "_meta": {
+  "source": "bt_10k",
+  "thresholds": "high:roi>10&n>=20 | low:roi<-5&n>=15 | unknown:n<10"
+ },
+ "lookup": {
+  "32n2r9bl6x90psj0wa7bfs6vq": {
+   "label": "high",
+   "bet_roi": 102.2,
+   "bet_n": 23,
+   "hit": 30.4,
+   "name": "Sudamericana"
+  },
+  "59tpnfrwnvhnhzmnvfyug68hj": {
+   "label": "high",
+   "bet_roi": 63.5,
+   "bet_n": 23,
+   "hit": 30.4,
+   "name": "Libertadores Kupası"
+  },
+  "b60nisd3qn427jm0hrg9kvmab": {
+   "label": "high",
+   "bet_roi": 49.7,
+   "bet_n": 22,
+   "hit": 22.7,
+   "name": "Allsvenskan"
+  },
+  "scf9p4y91yjvqvg5jndxzhxj": {
+   "label": "high",
+   "bet_roi": 33.8,
+   "bet_n": 100,
+   "hit": 25.0,
+   "name": "Serie A"
+  },
+  "4oogyu6o156iphvdvphwpck10": {
+   "label": "high",
+   "bet_roi": 32.3,
+   "bet_n": 23,
+   "hit": 21.7,
+   "name": "Şampiyonlar Ligi"
+  },
+  "89ovpy1rarewwzqvi30bfdr8b": {
+   "label": "high",
+   "bet_roi": 29.4,
+   "bet_n": 50,
+   "hit": 24.0,
+   "name": "1. Lig"
+  },
+  "82jkgccg7phfjpd0mltdl3pat": {
+   "label": "high",
+   "bet_roi": 25.8,
+   "bet_n": 29,
+   "hit": 27.6,
+   "name": "Süper Lig"
+  },
+  "3is4bkgf3loxv9qfg3hm8zfqb": {
+   "label": "high",
+   "bet_roi": 25.5,
+   "bet_n": 84,
+   "hit": 19.0,
+   "name": "LaLiga 2"
+  },
+  "enzlj1as2raqm4ids1zyb07y1": {
+   "label": "medium",
+   "bet_roi": 23.7,
+   "bet_n": 19,
+   "hit": 26.3,
+   "name": "USL 2. Lig"
+  },
+  "9ynnnx1qmkizq1o3qr3v0nsuk": {
+   "label": "high",
+   "bet_roi": 16.3,
+   "bet_n": 38,
+   "hit": 21.1,
+   "name": "Eliteserien"
+  },
+  "8ey0ww2zsosdmwr8ehsorh6t7": {
+   "label": "medium",
+   "bet_roi": 5.4,
+   "bet_n": 80,
+   "hit": 16.2,
+   "name": "Serie B"
+  },
+  "dm5ka0os1e3dxcp3vh05kmp33": {
+   "label": "low",
+   "bet_roi": -7.4,
+   "bet_n": 46,
+   "hit": 26.1,
+   "name": "Ligue 1"
+  },
+  "4zwgbb66rif2spcoeeol2motx": {
+   "label": "low",
+   "bet_roi": -12.7,
+   "bet_n": 39,
+   "hit": 23.1,
+   "name": "Pro Lig"
+  },
+  "a4fgj2rfbpf4ejo1qi624fefo": {
+   "label": "low",
+   "bet_roi": -14.2,
+   "bet_n": 73,
+   "hit": 17.8,
+   "name": "3. Lig"
+  },
+  "9chuiarcjofld1dkj9kysehmb": {
+   "label": "low",
+   "bet_roi": -14.9,
+   "bet_n": 22,
+   "hit": 13.6,
+   "name": "Superettan"
+  },
+  "3p81ltz6845appgkbgkzxueii": {
+   "label": "low",
+   "bet_roi": -19.8,
+   "bet_n": 34,
+   "hit": 14.7,
+   "name": "2. Lig"
+  },
+  "dvstmwnvw0mt5p38twn9yttyb": {
+   "label": "low",
+   "bet_roi": -37.2,
+   "bet_n": 19,
+   "hit": 26.3,
+   "name": "Veikkausliiga"
+  },
+  "zs18qaehvhg3w1208874zvfa": {
+   "label": "low",
+   "bet_roi": -62.0,
+   "bet_n": 17,
+   "hit": 23.5,
+   "name": "1. Lig"
+  }
+ }
+}

ai-engine/data/national_leagues.json

@@ -0,0 +1,31 @@
+{
+  "_meta": {
+    "purpose": "A-milli erkek futbol ligleri. betting_brain milli-maç gate'i bu listeyle tetiklenir.",
+    "strategy": "Milli maçta SADECE MS, oran 4.0-7.0, Hazırlık+Eleme oynanabilir. Turnuva/diğer market analiz-only. Backtest: +17% ROI, kararlılık-test geçti (eski/yeni yarı +22/+24%).",
+    "source": "2300-maç milli backtest (multi_backtest_20260602) segment+grid+stability analizi",
+    "competition_type_rule": "lig adı: 'hazırlık'->HAZIRLIK | 'eleme'/'play-off'->ELEME | diğer->TURNUVA"
+  },
+  "league_ids": [
+    "cesdwwnxbc5fmajgroc0hqzy2",
+    "3aa4mumjl6zyetg6o9hwd5hhx",
+    "40yjcbx2sq6oq736iqqqczwt1",
+    "39q1hq42hxjfylxb7xpe9bvf9",
+    "cu0rmpyff5692eo06ltddjo8a",
+    "ax1yf4nlzqpcji4j8epdgx3zl",
+    "1gxlzw2ezkyeykhcaa5x8ozkk",
+    "gfskxsdituog2kqp9yiu7bzi",
+    "595nsvo7ykvoe690b1e4u5n56",
+    "68zplepppndhl8bfdvgy9vgu1",
+    "3a0j0giz3c3ajw9h59evv7lqt",
+    "emy1ibc8fu2l0fukh4vlu5xl5",
+    "2db0aw1duj2my9l5iey5gm6nq",
+    "cc5tzz23tryrfqbm2pbv0jill",
+    "8tddm56zbasf57jkkay4kbf11",
+    "2r1hqz453bn9ljzt53kdr2lwb",
+    "93i7thp7zi0ympyt6l8aa1r2i",
+    "45db8orh1qttbsqq9hqapmbit",
+    "ude9t6yj60lebbn356qzg4k4",
+    "9qzn8cs96sgtqmesa9gpfti23",
+    "ad8y7vdjhinfqv4wo8rod6dck"
+  ]
+}

ai-engine/models/calibration.py

@@ -465,3 +465,105 @@ def get_calibrator() -> Calibrator:
     if _calibrator_instance is None:
         _calibrator_instance = Calibrator()
     return _calibrator_instance
+
+
+# ── FINAL-OUTPUT RECALIBRATION LAYER (V31e) ─────────────────────────────────
+# A thin, LAST-STEP per-market map: production calibrated_confidence -> reality.
+# Built from a 60-day backtest (scripts/fit_recalibrators.py); inference is a
+# pure np.interp over a 99-point monotone grid — NO sklearn needed at runtime.
+#
+# WHY THIS EXISTS:
+#   The upstream chain (temperature scaling T=1.5 -> per-outcome isotonic ->
+#   POST_CAL_TRUST blend) crushes high-base-rate binary markets toward 0.5,
+#   so "system says 51%" can really hit 70%. MS survives (near-uniform picks),
+#   which is why MS is already well-calibrated and OU/HT-OU markets are not.
+#
+# SAFETY / "DO NO HARM":
+#   * Only markets whose fit-time ECE >= 5.0 carry a map (currently OU15, OU35,
+#     HT_OU05, HT_OU15). MS and every already-good market have NO map ->
+#     recalibrate_conf() returns the input UNCHANGED -> guaranteed no regression.
+#   * Out-of-sample validated (fit=older 65%, test=unseen 35%):
+#       MS       ECE 1.1 -> 1.3  (flat, safe)
+#       HT_OU15      29.2 -> 0.8
+#       OU15         19.0 -> 3.3
+#       OU35         13.9 -> 4.3
+#       HT_OU05      11.5 -> 2.4
+#   * Adjusts ONLY the displayed confidence number. All rich analysis payload
+#     (probabilities, edges, vetoes, tiers, bands) is preserved untouched, and
+#     the pre-recalibration value is kept for audit by the caller.
+FINAL_RECALIBRATOR_PATH = os.path.join(CALIBRATION_DIR, "final_recalibrators.json")
+
+
+class FinalRecalibrator:
+    """Per-market final-output recalibration via piecewise-linear interpolation.
+
+    Loads a compact JSON of 99-point lookup grids (x=calibrated_confidence/100,
+    y=reality). Markets absent from the file pass through as identity.
+    """
+
+    def __init__(self, path: str = FINAL_RECALIBRATOR_PATH):
+        self.grid: Optional[np.ndarray] = None
+        self.maps: Dict[str, np.ndarray] = {}
+        self.source_path = path
+        self._load(path)
+
+    def _load(self, path: str) -> None:
+        if not os.path.exists(path):
+            print(f"[FinalRecalibrator] No map file at {path} — pass-through mode (all markets unchanged)")
+            return
+        try:
+            with open(path, "r") as f:
+                data = json.load(f)
+            meta = data.get("_meta", {})
+            grid = meta.get("grid")
+            if not grid:
+                print("[FinalRecalibrator] Map file missing _meta.grid — pass-through mode")
+                return
+            self.grid = np.asarray(grid, dtype=float)
+            for market, m in data.items():
+                if market == "_meta" or not isinstance(m, dict):
+                    continue
+                y = m.get("y")
+                if y and len(y) == len(self.grid):
+                    self.maps[str(market).upper()] = np.asarray(y, dtype=float)
+                else:
+                    print(f"[FinalRecalibrator] Skipped {market}: grid/y length mismatch")
+            print(f"[FinalRecalibrator] Loaded reality maps for {sorted(self.maps.keys())} "
+                  f"(everything else, incl. MS, passes through unchanged)")
+        except Exception as e:
+            print(f"[FinalRecalibrator] Warning: failed to load {path}: {e} — pass-through mode")
+            self.grid = None
+            self.maps = {}
+
+    def has_map(self, market: str) -> bool:
+        return bool(self.maps) and (market or "").upper() in self.maps
+
+    def recalibrate_conf(self, market: str, calibrated_conf: float) -> float:
+        """Map a 0–100 confidence to its reality-aligned value.
+
+        Markets without a trained map (including MS and all already-good
+        markets) return the input UNCHANGED. Any failure also returns the
+        input unchanged so this layer can never regress production.
+        """
+        try:
+            key = (market or "").upper()
+            if self.grid is None or key not in self.maps:
+                return calibrated_conf
+            x = float(calibrated_conf) / 100.0
+            x = min(max(x, 0.0), 1.0)
+            y = float(np.interp(x, self.grid, self.maps[key]))
+            return max(1.0, min(99.0, y * 100.0))
+        except Exception:
+            return calibrated_conf
+
+
+# Singleton instance
+_final_recalibrator_instance: Optional[FinalRecalibrator] = None
+
+
+def get_final_recalibrator() -> FinalRecalibrator:
+    """Get or create the global FinalRecalibrator instance."""
+    global _final_recalibrator_instance
+    if _final_recalibrator_instance is None:
+        _final_recalibrator_instance = FinalRecalibrator()
+    return _final_recalibrator_instance

ai-engine/models/calibration/final_recalibrators.json

@@ -0,0 +1,532 @@
+{
+ "_meta": {
+  "grid": [
+   0.01,
+   0.02,
+   0.03,
+   0.04,
+   0.05,
+   0.06,
+   0.07,
+   0.08,
+   0.09,
+   0.1,
+   0.11,
+   0.12,
+   0.13,
+   0.14,
+   0.15,
+   0.16,
+   0.17,
+   0.18,
+   0.19,
+   0.2,
+   0.21,
+   0.22,
+   0.23,
+   0.24,
+   0.25,
+   0.26,
+   0.27,
+   0.28,
+   0.29,
+   0.3,
+   0.31,
+   0.32,
+   0.33,
+   0.34,
+   0.35,
+   0.36,
+   0.37,
+   0.38,
+   0.39,
+   0.4,
+   0.41,
+   0.42,
+   0.43,
+   0.44,
+   0.45,
+   0.46,
+   0.47,
+   0.48,
+   0.49,
+   0.5,
+   0.51,
+   0.52,
+   0.53,
+   0.54,
+   0.55,
+   0.56,
+   0.57,
+   0.58,
+   0.59,
+   0.6,
+   0.61,
+   0.62,
+   0.63,
+   0.64,
+   0.65,
+   0.66,
+   0.67,
+   0.68,
+   0.69,
+   0.7,
+   0.71,
+   0.72,
+   0.73,
+   0.74,
+   0.75,
+   0.76,
+   0.77,
+   0.78,
+   0.79,
+   0.8,
+   0.81,
+   0.82,
+   0.83,
+   0.84,
+   0.85,
+   0.86,
+   0.87,
+   0.88,
+   0.89,
+   0.9,
+   0.91,
+   0.92,
+   0.93,
+   0.94,
+   0.95,
+   0.96,
+   0.97,
+   0.98,
+   0.99
+  ],
+  "threshold_ece": 5.0,
+  "source": "/tmp/multi_60d.csv",
+  "note": "x=calibrated_confidence/100; new=interp(grid,y)"
+ },
+ "HT_OU05": {
+  "grid_min": 0.01,
+  "grid_max": 0.99,
+  "n": 3683,
+  "y": [
+   0.0833,
+   0.3333,
+   0.3333,
+   0.3333,
+   0.3394,
+   0.3636,
+   0.3636,
+   0.3636,
+   0.3636,
+   0.3636,
+   0.3636,
+   0.3636,
+   0.3636,
+   0.3727,
+   0.3955,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4,
+   0.4583,
+   0.6286,
+   0.6286,
+   0.6286,
+   0.6286,
+   0.6286,
+   0.6286,
+   0.6286,
+   0.6286,
+   0.6286,
+   0.6531,
+   0.672,
+   0.7143,
+   0.7262,
+   0.7262,
+   0.7312,
+   0.7406,
+   0.7655,
+   0.7655,
+   0.8495,
+   0.8495,
+   0.8495,
+   0.8495,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0
+  ]
+ },
+ "HT_OU15": {
+  "grid_min": 0.01,
+  "grid_max": 0.99,
+  "n": 5200,
+  "y": [
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4118,
+   0.4521,
+   0.5385,
+   0.5385,
+   0.5385,
+   0.5848,
+   0.6142,
+   0.6142,
+   0.6142,
+   0.6245,
+   0.6245,
+   0.6245,
+   0.6262,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6275,
+   0.6452,
+   0.6842,
+   0.6842,
+   0.6842,
+   0.6842,
+   0.6842,
+   0.6842,
+   0.8077,
+   0.8077,
+   0.8077,
+   0.8077,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0
+  ]
+ },
+ "OU15": {
+  "grid_min": 0.01,
+  "grid_max": 0.99,
+  "n": 2724,
+  "y": [
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.2797,
+   0.4352,
+   0.6295,
+   0.7165,
+   0.7174,
+   0.7987,
+   0.8197,
+   0.8197,
+   0.8197,
+   0.8197,
+   0.8197,
+   0.8197,
+   0.9118,
+   0.9276,
+   0.9502,
+   0.9729,
+   0.9955,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0,
+   1.0
+  ]
+ },
+ "OU35": {
+  "grid_min": 0.01,
+  "grid_max": 0.99,
+  "n": 4277,
+  "y": [
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.0,
+   0.474,
+   0.474,
+   0.474,
+   0.474,
+   0.474,
+   0.474,
+   0.474,
+   0.474,
+   0.474,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.571,
+   0.6222,
+   0.6222,
+   0.6222,
+   0.6222,
+   0.6222,
+   0.7747,
+   0.7747,
+   0.7747,
+   0.7747,
+   0.7747,
+   0.7788,
+   0.8195,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.8333,
+   0.836,
+   0.8624,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889,
+   0.8889
+  ]
+ }
+}

ai-engine/models/live_matrix.py

@@ -0,0 +1,184 @@
+"""Live-conditioned score projection (V38) — pure functions, no I/O.
+
+Answers, DURING a match, questions like "1-0 at 80' — what is the REAL
+probability the away team still scores?" by conditioning the same calibrated
+market-anchored lambdas (V35/V36) on the current score and minute.
+
+Mechanics — a minute-stepped Markov chain over remaining goals:
+
+  1. Pre-match lambdas come from the SAME source the score card uses
+     (de-vigged 1X2 + over2.5, models/score_matrix solvers) — one consistent
+     probability spine pre-match and in-play.
+  2. Each remaining minute contributes lambda_side x minute_share(t) goals,
+     where minute_share is the EMPIRICAL goal-time intensity curve measured
+     on 38,779 clean-timeline real-odds matches (1H share 44.4%, late-game
+     intensity rises, stoppage spikes at 45' and 90+').
+  3. Each minute's intensity is scaled by the MEASURED score-state
+     multiplier: trailing teams push (+9%, +17% after 70'), leading teams
+     shut up shop (-5%/-7%), 2+ ahead opens up. The chain updates the state
+     as virtual goals happen, so multipliers switch mid-projection exactly
+     like they do on the pitch.
+
+All constants are fitted on the train window (matches older than the last 90
+days); the held-out window validates calibration out-of-sample before any of
+this reaches the screen.
+"""
+
+from __future__ import annotations
+
+from typing import Dict, List, Optional, Tuple
+
+from models.score_matrix import split_lambdas, total_lambda_from_over25
+
+MAX_MINUTE = 94          # 90 + folded stoppage
+LATE_PHASE_FROM = 70     # measured multipliers switch here
+MAX_EXTRA_GOALS = 7      # per side, absorbing cap for the chain
+
+# Empirical goal-time intensity: share of a match's goals per 5-min bucket
+# (0-5, ..., 90-94+). Measured on 105k goals; 45' and 90+' buckets carry the
+# folded stoppage-time spikes.
+INTENSITY_SHARES: Tuple[float, ...] = (
+    0.036, 0.045, 0.047, 0.047, 0.045, 0.046, 0.048, 0.049, 0.081,
+    0.048, 0.057, 0.055, 0.054, 0.053, 0.052, 0.053, 0.052, 0.056, 0.076,
+)
+
+# Score-state goal-intensity multipliers, measured (actual/expected) by the
+# scoring side's goal difference, split early (<70') / late (>=70').
+_STATE_MULT_EARLY: Dict[int, float] = {-2: 1.095, -1: 1.045, 0: 0.966, 1: 0.952, 2: 1.011}
+_STATE_MULT_LATE: Dict[int, float] = {-2: 1.123, -1: 1.174, 0: 1.015, 1: 0.930, 2: 1.011}
+
+
+def _minute_share(minute: int) -> float:
+    """Per-minute share of match-total goal intensity at `minute` (1-based)."""
+    b = min(len(INTENSITY_SHARES) - 1, max(0, (minute - 1) // 5))
+    return INTENSITY_SHARES[b] / 5.0
+
+
+def state_multiplier(diff: int, minute: int) -> float:
+    """Intensity multiplier for a side whose current goal difference is
+    `diff` (own − opponent), at `minute`."""
+    d = max(-2, min(2, diff))
+    table = _STATE_MULT_LATE if minute >= LATE_PHASE_FROM else _STATE_MULT_EARLY
+    return table[d]
+
+
+def estimate_minute(match_date_ms: Optional[int], now_ms: int) -> Optional[int]:
+    """Approximate current match minute from kickoff time (no feed minute is
+    available: live_matches.substate carries none). Folds the ~15' half-time
+    break; accuracy is ±2-3 minutes which barely moves the projection."""
+    if not match_date_ms:
+        return None
+    elapsed = (now_ms - int(match_date_ms)) / 60000.0
+    if elapsed < 0:
+        return None
+    if elapsed <= 48:                      # first half (+stoppage)
+        minute = elapsed
+    elif elapsed <= 63:                    # half-time break window
+        minute = 46
+    else:
+        minute = elapsed - 15.0            # second half, break folded out
+    return int(max(1, min(MAX_MINUTE, minute)))
+
+
+def _chain(
+    lam_h: float,
+    lam_a: float,
+    cur_h: int,
+    cur_a: int,
+    minute: int,
+) -> Dict[Tuple[int, int], float]:
+    """Distribution over (extra home goals, extra away goals) from `minute`
+    to full time, with state-dependent intensities."""
+    dist: Dict[Tuple[int, int], float] = {(0, 0): 1.0}
+    for t in range(minute, MAX_MINUTE + 1):
+        share = _minute_share(t)
+        nxt: Dict[Tuple[int, int], float] = {}
+        for (eh, ea), p in dist.items():
+            diff = (cur_h + eh) - (cur_a + ea)
+            ph = lam_h * share * state_multiplier(diff, t)
+            pa = lam_a * share * state_multiplier(-diff, t)
+            ph = min(ph, 0.30); pa = min(pa, 0.30)
+            stay = max(0.0, 1.0 - ph - pa)
+            nxt[(eh, ea)] = nxt.get((eh, ea), 0.0) + p * stay
+            if eh < MAX_EXTRA_GOALS:
+                nxt[(eh + 1, ea)] = nxt.get((eh + 1, ea), 0.0) + p * ph
+            else:
+                nxt[(eh, ea)] = nxt.get((eh, ea), 0.0) + p * ph
+            if ea < MAX_EXTRA_GOALS:
+                nxt[(eh, ea + 1)] = nxt.get((eh, ea + 1), 0.0) + p * pa
+            else:
+                nxt[(eh, ea)] = nxt.get((eh, ea), 0.0) + p * pa
+        dist = nxt
+    return dist
+
+
+def build_live_projection(
+    p1: float,
+    px: float,
+    p2: float,
+    p_over25: float,
+    cur_h: int,
+    cur_a: int,
+    minute: int,
+) -> Dict[str, object]:
+    """Live projection from the anchored pre-match probabilities + the pitch
+    state. Returns honest, score/minute-aware probabilities.
+
+    (p1, px, p2) and p_over25 are the CALIBRATED (V35-anchored) numbers; the
+    same spine the pre-match cards display.
+    """
+    minute = int(max(1, min(MAX_MINUTE, minute)))
+    cur_h = max(0, int(cur_h)); cur_a = max(0, int(cur_a))
+    total = total_lambda_from_over25(p_over25)
+    lam_h, lam_a = split_lambdas(total, p1, p2)
+
+    dist = _chain(lam_h, lam_a, cur_h, cur_a, minute)
+
+    p_home_win = p_draw = p_away_win = 0.0
+    p_home_scores = p_away_scores = 0.0
+    exp_goals = 0.0
+    scores: Dict[str, float] = {}
+    for (eh, ea), p in dist.items():
+        fh, fa = cur_h + eh, cur_a + ea
+        if fh > fa: p_home_win += p
+        elif fh == fa: p_draw += p
+        else: p_away_win += p
+        if eh > 0: p_home_scores += p
+        if ea > 0: p_away_scores += p
+        exp_goals += p * (eh + ea)
+        key = f"{min(fh,9)}-{min(fa,9)}"
+        scores[key] = scores.get(key, 0.0) + p
+
+    top = sorted(scores.items(), key=lambda kv: kv[1], reverse=True)[:5]
+    total_now = cur_h + cur_a
+    p_over25_live = sum(
+        p for (eh, ea), p in dist.items() if total_now + eh + ea >= 3
+    )
+
+    # "comeback": the side currently behind at least draws / currently level
+    # match does NOT stay level
+    if cur_h > cur_a:
+        p_comeback = p_draw + p_away_win
+    elif cur_a > cur_h:
+        p_comeback = p_draw + p_home_win
+    else:
+        p_comeback = p_home_win + p_away_win  # deadlock breaks
+
+    return {
+        "minute": minute,
+        "current_score": f"{cur_h}-{cur_a}",
+        "probs": {
+            "1": round(p_home_win, 4),
+            "X": round(p_draw, 4),
+            "2": round(p_away_win, 4),
+        },
+        "p_home_scores_again": round(p_home_scores, 4),
+        "p_away_scores_again": round(p_away_scores, 4),
+        "p_comeback": round(p_comeback, 4),
+        "p_over25": round(p_over25_live, 4),
+        "expected_remaining_goals": round(exp_goals, 2),
+        "scenario_top5": [
+            {"score": s, "prob": round(p, 4)} for s, p in top
+        ],
+        "calibration_source": "live_matrix_v38",
+    }

ai-engine/models/market_anchor.py

@@ -0,0 +1,258 @@
+"""Market-anchored calibration (V35) — pure functions, no I/O.
+
+WHY THIS EXISTS
+---------------
+The model's invented per-market probabilities were *measured* to be badly
+overconfident. Grading the engine's own stored predictions against actual
+results: it says ~50% where reality is ~25%, ~67% where reality is ~37%
+(calibration error / ECE on the order of 25-30%). That mis-calibration is the
+direct cause of the false "value" picks and the negative realised ROI.
+
+The de-vigged market price, by contrast, is empirically near-perfectly
+calibrated. Out-of-sample (correction fit on 2023-24, tested on 2025-26;
+78k real-odds football matches) the de-vigged market's ECE was:
+    home 1.56% | draw 1.85% | away 1.49% | over2.5 1.79% | btts 1.38%
+Adding one small, large-sample home-favourite correction cut MS-home ECE
+from 1.56% -> 0.64%.
+
+So for the DISPLAYED probabilities we anchor to the de-vigged market and apply
+only that one proven correction. ~20-40x more calibrated than the model's
+numbers, and fully transparent.
+
+These functions are pure (stdlib only) so they can be unit-tested in isolation
+without the DB or the heavy model stack.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import threading
+import time
+from typing import Any, Dict, List, Optional, Tuple
+
+
+def devig(odds: List[Optional[float]]) -> Optional[List[float]]:
+    """Vig-removed (fair) probabilities from a group of decimal odds.
+
+    ``p_i = (1/odds_i) / Σ(1/odds_j)`` — normalising the raw implied
+    probabilities to sum to 1 removes the bookmaker margin.
+
+    Returns ``None`` when ANY leg is missing or non-real (``<= 1.01``). That is
+    deliberate: a market with a missing/placeholder leg has no real price, and
+    the product rule is to never fabricate numbers for a match without odds.
+    """
+    if not odds or any(o is None or float(o) <= 1.01 for o in odds):
+        return None
+    inv = [1.0 / float(o) for o in odds]
+    total = sum(inv)
+    if total <= 0.0:
+        return None
+    return [x / total for x in inv]
+
+
+# Home-favourite correction: measured (actual home-win rate − de-vigged implied)
+# by implied-home band, out-of-sample on real-odds matches. Big home favourites
+# win a few points MORE than the de-vigged price implies; underdogs are roughly
+# unbiased. Values are deliberately conservative — universal and shrunk toward 0
+# vs the raw tier-0 (soft-league) edge, because the bias is weaker in efficient
+# top leagues. Applying these took MS-home OOS ECE 1.56% -> 0.64%.
+#
+# These static bands are the BUILT-IN FALLBACK. The live values come from the
+# versioned artifact `config/market_anchor_corrections.json`, refreshed by
+# `scripts/fit_anchor_corrections.py` (the guarded self-correction loop:
+# measure on settled matches -> shrink/clip/min-sample gates -> out-of-sample
+# acceptance -> write table). The engine only ever consumes the TABLE — the
+# loop never modifies code.
+_HOME_FAV_BANDS: Tuple[Tuple[float, float, float], ...] = (
+    (0.45, 0.55, 0.010),
+    (0.55, 0.65, 0.018),
+    (0.65, 0.75, 0.028),
+    (0.75, 1.01, 0.034),
+)
+
+_DEFAULT_CORRECTIONS_PATH = os.path.join(
+    os.path.dirname(os.path.abspath(__file__)), "..", "config",
+    "market_anchor_corrections.json",
+)
+
+
+def _corrections_path() -> str:
+    return os.environ.get(
+        "MARKET_ANCHOR_CORRECTIONS_PATH", _DEFAULT_CORRECTIONS_PATH
+    )
+_corrections_lock = threading.Lock()
+_corrections_cache: Optional[Dict[str, Any]] = None
+_corrections_ts: float = 0.0
+# Re-check sources at most every 10 minutes: the self-correction cron writes a
+# new table to app_settings; running engines pick it up WITHOUT a restart.
+_CORRECTIONS_TTL_S = 600.0
+
+
+def _parse_corrections(raw: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+    parsed_table: Dict[str, Any] = {}
+    for key in ("ms_home", "ms_away"):
+        bands = raw.get("corrections", {}).get(key)
+        if not (isinstance(bands, list) and bands):
+            continue
+        parsed = []
+        for b in bands:
+            lo = float(b["lo"]); hi = float(b["hi"]); delta = float(b["delta"])
+            if not (0.0 <= lo < hi <= 1.01) or abs(delta) > 0.10:
+                raise ValueError(f"correction band out of range: {b}")
+            parsed.append((lo, hi, delta))
+        parsed_table[key] = tuple(parsed)
+    if not parsed_table:
+        return None
+    parsed_table["version"] = str(raw.get("version", "?"))
+    return parsed_table
+
+
+def _db_corrections_raw() -> Optional[Dict[str, Any]]:
+    """Fetch the correction artifact from app_settings (the deployment's shared
+    medium — the ai-engine container has no volume mounts, so a host-side cron
+    can only reach the running engine through the database). Guarded: any
+    failure → None, never breaks a prediction. Disable with MARKET_ANCHOR_DB=0."""
+    if os.environ.get("MARKET_ANCHOR_DB", "1") == "0":
+        return None
+    try:
+        import psycopg2  # local import: keeps module usable without DB deps
+        from data.db import get_clean_dsn
+
+        with psycopg2.connect(get_clean_dsn(), connect_timeout=3) as conn:
+            with conn.cursor() as cur:
+                cur.execute(
+                    "SELECT value FROM app_settings"
+                    " WHERE key = 'market_anchor_corrections'"
+                )
+                row = cur.fetchone()
+        if row and row[0]:
+            return json.loads(row[0])
+    except Exception:
+        return None
+    return None
+
+
+def _load_corrections() -> Optional[Dict[str, Any]]:
+    """Resolve the active correction table (thread-safe, TTL-cached).
+
+    Source order:
+      1. MARKET_ANCHOR_CORRECTIONS_PATH env file (tests/dev — file-only mode,
+         malformed → static fallback, DB and default file are NOT consulted)
+      2. app_settings DB row 'market_anchor_corrections' (production path —
+         refreshed by scripts/fit_anchor_corrections.py)
+      3. bundled config/market_anchor_corrections.json
+      4. None → built-in static fallback bands
+    """
+    global _corrections_cache, _corrections_ts
+    now = time.time()
+    if now - _corrections_ts < _CORRECTIONS_TTL_S:
+        return _corrections_cache
+    with _corrections_lock:
+        if now - _corrections_ts < _CORRECTIONS_TTL_S:
+            return _corrections_cache
+        table: Optional[Dict[str, Any]] = None
+        env_path = os.environ.get("MARKET_ANCHOR_CORRECTIONS_PATH")
+        if env_path:
+            try:
+                with open(env_path, "r", encoding="utf-8") as fh:
+                    table = _parse_corrections(json.load(fh))
+            except (OSError, ValueError, KeyError, TypeError, json.JSONDecodeError):
+                table = None
+        else:
+            raw = _db_corrections_raw()
+            if raw is not None:
+                try:
+                    table = _parse_corrections(raw)
+                except (ValueError, KeyError, TypeError):
+                    table = None
+            if table is None:
+                try:
+                    with open(_corrections_path(), "r", encoding="utf-8") as fh:
+                        table = _parse_corrections(json.load(fh))
+                except (OSError, ValueError, KeyError, TypeError, json.JSONDecodeError):
+                    table = None
+        _corrections_cache = table
+        _corrections_ts = time.time()
+        return _corrections_cache
+
+
+def reload_corrections() -> None:
+    """Force re-read of the correction sources (used after a refresh/tests)."""
+    global _corrections_ts, _corrections_cache
+    with _corrections_lock:
+        _corrections_ts = 0.0
+        _corrections_cache = None
+
+
+def home_favorite_delta(p_home: float) -> float:
+    """Additive correction to the de-vigged home-win probability.
+
+    Band semantics: a fitted-artifact band OVERRIDES the static prior where it
+    exists (including an explicit delta of 0 — evidence of "no bias"). Where
+    the artifact is SILENT (a range that never passed the min-sample gate,
+    e.g. big favourites 0.75+), the static prior still applies — missing
+    evidence must not silently erase proven knowledge."""
+    table = _load_corrections()
+    if table and "ms_home" in table:
+        for lo, hi, delta in table["ms_home"]:
+            if lo <= p_home < hi:
+                return delta
+    for lo, hi, delta in _HOME_FAV_BANDS:
+        if lo <= p_home < hi:
+            return delta
+    return 0.0
+
+
+def away_favorite_delta(p_away: float) -> float:
+    """Additive correction to the de-vigged away-win probability.
+
+    Scoreboard measurement (2026-06): away favourites also win a few points
+    MORE than the de-vigged price implies (+2.6..+4.2pt). Unlike the home
+    side there is NO built-in fallback — away corrections must be EARNED via
+    the fitted artifact (scripts/fit_anchor_corrections.py passing its
+    out-of-sample acceptance gate). No artifact → zero → prior behaviour."""
+    table = _load_corrections()
+    bands = table.get("ms_away", ()) if table else ()
+    for lo, hi, delta in bands:
+        if lo <= p_away < hi:
+            return delta
+    return 0.0
+
+
+def apply_corrections(
+    p1: float, px: float, p2: float
+) -> Tuple[float, float, float]:
+    """Apply favourite corrections to a 3-way (1, X, 2) vector.
+
+    In practice only one side can be a favourite (both ≥0.45 would leave no
+    room for the draw); if both bands somehow fire, the larger delta wins.
+    The other two outcomes are renormalised so the vector still sums to 1."""
+    d1 = home_favorite_delta(p1)
+    d2 = away_favorite_delta(p2)
+    if d1 <= 0.0 and d2 <= 0.0:
+        return p1, px, p2
+    if d1 >= d2:
+        return apply_home_correction(p1, px, p2)
+    p2n = min(0.98, p2 + d2)
+    remaining = 1.0 - p2n
+    rest = p1 + px
+    if rest <= 0.0:
+        return p1, px, p2n
+    return p1 / rest * remaining, px / rest * remaining, p2n
+
+
+def apply_home_correction(
+    p1: float, px: float, p2: float
+) -> Tuple[float, float, float]:
+    """Apply the home-favourite delta to a 3-way (1, X, 2) probability vector,
+    renormalising draw/away so the three still sum to 1.0."""
+    delta = home_favorite_delta(p1)
+    if delta <= 0.0:
+        return p1, px, p2
+    p1n = min(0.98, p1 + delta)
+    remaining = 1.0 - p1n
+    rest = px + p2
+    if rest <= 0.0:
+        return p1n, px, p2
+    return p1n, px / rest * remaining, p2 / rest * remaining

ai-engine/models/score_matrix.py

@@ -0,0 +1,166 @@
+"""Market-anchored score matrix (V36) — pure functions, no I/O.
+
+WHY THIS EXISTS
+---------------
+The engine's displayed score predictions (`score_prediction`, `scenario_top5`)
+come from the model's invented xG, so they can contradict the calibrated
+market-anchored probabilities shown right next to them (V35). Example seen in
+production: MS card says home 78% while the score card's distribution implies
+something else entirely.
+
+This module derives the FULL scoreline distribution from the SAME calibrated
+(de-vigged) market probabilities that the V35 market anchor displays:
+
+    1. Solve total-goals lambda T from the calibrated P(over 2.5)
+       (total goals ~ Poisson(T):  P(N>=3) = 1 - e^-T (1 + T + T^2/2)).
+    2. Split T into (lambda_home, lambda_away) so the independent-Poisson
+       matrix's home/away win gap matches the calibrated 1X2.
+    3. Build the score matrix, then IPF-scale the three outcome regions
+       (home-win cells, draw cells, away-win cells) so they sum EXACTLY to the
+       calibrated (p1, px, pX2) — guaranteeing the score card and the MS card
+       can never disagree again.
+    4. Half-time matrix: same machinery with lambdas scaled by the measured
+       first-half goal share, optionally IPF'd to the anchored HT 1X2.
+
+All stdlib (math only) → unit-testable in isolation, no model/DB deps.
+
+Validated on 63,681 real-odds matches (2025-26, out-of-sample constants):
+see tests + the calibration session notes. Honest ceiling reminder: even a
+perfect correct-score predictor only hits the modal score ~12-15% of the time;
+the value here is honest, consistent probabilities — not certainty.
+"""
+
+from __future__ import annotations
+
+import math
+from typing import Dict, List, Optional, Tuple
+
+# Measured on 63,681 real-odds matches (2025-26): share of full-time goals
+# scored in the first half, per side (home 0.4440, away 0.4428).
+HT_GOAL_SHARE_HOME = 0.44
+HT_GOAL_SHARE_AWAY = 0.44
+
+MAX_GOALS = 10  # matrix is (0..10)x(0..10); tail mass beyond is negligible
+
+
+def _pois_pmf(lam: float, k: int) -> float:
+    return math.exp(-lam) * lam**k / math.factorial(k)
+
+
+def total_lambda_from_over25(p_over25: float) -> float:
+    """Solve T such that P(Poisson(T) >= 3) == p_over25, by bisection."""
+    p = min(max(p_over25, 0.01), 0.99)
+
+    def p_over(t: float) -> float:
+        return 1.0 - math.exp(-t) * (1.0 + t + t * t / 2.0)
+
+    lo, hi = 0.05, 8.0
+    for _ in range(60):
+        mid = (lo + hi) / 2.0
+        if p_over(mid) < p:
+            lo = mid
+        else:
+            hi = mid
+    return (lo + hi) / 2.0
+
+
+def _raw_matrix(lh: float, la: float) -> List[List[float]]:
+    ph = [_pois_pmf(lh, i) for i in range(MAX_GOALS + 1)]
+    pa = [_pois_pmf(la, j) for j in range(MAX_GOALS + 1)]
+    return [[ph[i] * pa[j] for j in range(MAX_GOALS + 1)] for i in range(MAX_GOALS + 1)]
+
+
+def _outcome_sums(mat: List[List[float]]) -> Tuple[float, float, float]:
+    w = d = l = 0.0
+    for i in range(MAX_GOALS + 1):
+        for j in range(MAX_GOALS + 1):
+            if i > j:
+                w += mat[i][j]
+            elif i == j:
+                d += mat[i][j]
+            else:
+                l += mat[i][j]
+    return w, d, l
+
+
+def split_lambdas(total: float, p1: float, p2: float) -> Tuple[float, float]:
+    """Split total lambda into (home, away) so the matrix's win-prob gap
+    matches the calibrated 1X2 gap, by bisection on the home share."""
+    target_gap = p1 - p2
+    lo, hi = 0.10, 0.90
+    for _ in range(40):
+        s = (lo + hi) / 2.0
+        w, _, l = _outcome_sums(_raw_matrix(total * s, total * (1.0 - s)))
+        if (w - l) < target_gap:
+            lo = s
+        else:
+            hi = s
+    s = (lo + hi) / 2.0
+    return total * s, total * (1.0 - s)
+
+
+def ipf_to_outcomes(
+    mat: List[List[float]], p1: float, px: float, p2: float
+) -> List[List[float]]:
+    """Scale the home-win / draw / away-win regions so each sums EXACTLY to the
+    calibrated (p1, px, p2). This is what makes the score card mathematically
+    consistent with the displayed MS probabilities."""
+    w, d, l = _outcome_sums(mat)
+    if min(w, d, l) <= 0.0:
+        return mat
+    fw, fd, fl = p1 / w, px / d, p2 / l
+    out = [[0.0] * (MAX_GOALS + 1) for _ in range(MAX_GOALS + 1)]
+    for i in range(MAX_GOALS + 1):
+        for j in range(MAX_GOALS + 1):
+            f = fw if i > j else fd if i == j else fl
+            out[i][j] = mat[i][j] * f
+    return out
+
+
+def top_scores(mat: List[List[float]], n: int = 5) -> List[Dict[str, object]]:
+    cells = [
+        (mat[i][j], i, j)
+        for i in range(MAX_GOALS + 1)
+        for j in range(MAX_GOALS + 1)
+    ]
+    cells.sort(reverse=True)
+    return [
+        {"score": f"{i}-{j}", "prob": round(p, 4)}
+        for p, i, j in cells[:n]
+    ]
+
+
+def build_calibrated_score_package(
+    p1: float,
+    px: float,
+    p2: float,
+    p_over25: float,
+    ht_probs: Optional[Tuple[float, float, float]] = None,
+) -> Dict[str, object]:
+    """Full calibrated score card from the V35-anchored probabilities.
+
+    Returns {ft, ht, xg_home, xg_away, xg_total, scenario_top5, ht_top}.
+    xg_* here are MARKET-implied goal expectations (the lambdas), so every
+    number on the card comes from one consistent source.
+    """
+    total = total_lambda_from_over25(p_over25)
+    lh, la = split_lambdas(total, p1, p2)
+    ft_mat = ipf_to_outcomes(_raw_matrix(lh, la), p1, px, p2)
+    ft_top = top_scores(ft_mat, 5)
+
+    lh_ht, la_ht = lh * HT_GOAL_SHARE_HOME, la * HT_GOAL_SHARE_AWAY
+    ht_mat = _raw_matrix(lh_ht, la_ht)
+    if ht_probs is not None:
+        ht_mat = ipf_to_outcomes(ht_mat, *ht_probs)
+    ht_top = top_scores(ht_mat, 3)
+
+    return {
+        "ft": str(ft_top[0]["score"]) if ft_top else None,
+        "ht": str(ht_top[0]["score"]) if ht_top else None,
+        "xg_home": round(lh, 2),
+        "xg_away": round(la, 2),
+        "xg_total": round(lh + la, 2),
+        "scenario_top5": ft_top,
+        "ht_top": ht_top,
+        "calibration_source": "market_anchor_v36_score",
+    }

ai-engine/reports/V29_OPTIMIZATION_REPORT.md

@@ -0,0 +1,75 @@
+# V29 Data-Driven Optimization Report
+## Based on 7,000-match Diagnostic Backtest (2026-05-27)
+
+### Before (V28-Pro-Max)
+- **4,134 settled BET-action picks**
+- Hit rate: 54.9%
+- Unit profit: -132.68
+- Staked: 849.50
+- **ROI: -15.6%**
+
+### Root Cause Analysis
+
+#### 1. Value Sniper Threshold Too Loose (CRITICAL)
+```python
+# OLD: ev_edge >= 0.008 or calibrated_conf >= 55.0
+# This made 100% of bets qualify as "value sniper", bypassing ALL betting brain vetoes
+```
+- 4,134/4,134 bets (100%) had `is_value_sniper = True`
+- Hard vetoes (negative_ev, market_muted, low_reliability) were NEVER enforced
+
+#### 2. 89% of Bets Had Negative EV Edge
+- n=3,688 with ev_edge < 0: ROI = -16.1%
+- The model was systematically pricing below market, meaning every bet carried negative expected value
+
+#### 3. OU25 Market Unprofitable in ALL Configurations
+- n=1,563 bets, -17.1% ROI
+- Even with ev>=5% + rel>=0.55: n=27, -36.9% ROI
+- Grid search found NO profitable filter combination
+
+#### 4. BTTS Market Marginal
+- n=1,456 bets, -15.4% ROI
+- Only profitable with ev>=5%: n=15, +12.9% (but tiny sample)
+
+### Grid Search Results (Top Profitable Combos)
+
+| Market | EV Min | Rel Min | V27 | n | Hit% | ROI |
+|--------|--------|---------|-----|---|------|-----|
+| MS | >=5% | >=0.55 | AGREE | 42 | 59.5% | **+10.4%** |
+| MS | >=5% | >=0.55 | ANY | 52 | 59.6% | **+8.6%** |
+| MS | >=3% | >=0.55 | ANY | 69 | 56.5% | **+4.0%** |
+| BTTS | >=5% | >=0.70 | ANY | 15 | 60.0% | **+12.9%** |
+| MS | >=5% | >=0.00 | ANY | 113 | 55.8% | **-0.7%** |
+
+### Changes Applied (V29)
+
+#### market_board.py
+```python
+# Tightened from: ev >= 0.008 OR conf >= 55.0
+# To: ALL three must be true
+is_value_sniper = ev_edge >= 0.05 and calibrated_conf >= 60.0 and odds_rel >= 0.55
+```
+
+#### betting_brain.py
+1. **MIN_BET_SCORE**: 72.0 -> 62.0 (hard vetoes now do the filtering)
+2. **MIN_WATCH_SCORE**: 62.0 -> 52.0
+3. **MUTED_MARKETS**: `{"BTTS"}` -> `{"OU25", "DC", "OU35"}`
+4. **MARKET_OPTIMAL_FILTERS**:
+   - MS: min_edge=0.03, min_reliability=0.55, require_v27_agree=False
+   - BTTS: min_edge=0.05, min_reliability=0.70 (strict envelope)
+5. **Hard vetoes no longer bypassed by sniper**:
+   - `negative_ev_edge` (ev < 0)
+   - `ev_edge_too_high_trap` (ev >= 0.20)
+   - `market_muted_by_backtest`
+   - `low_reliability_league_hard_block` (rel < 0.30)
+   - Per-market envelope checks
+
+### Expected Performance (Simulated on 7K backtest)
+- **65 bets** out of 7,000 matches (0.9% selectivity)
+- Hit rate: 56.9%
+- **ROI: +6.8%** (from -15.6%)
+- MS dominates: n=64, ROI=+8.0%
+- Consistent: April +14.0%, May +4.9%
+
+### Trade-off
+The system becomes very selective (fewer bets per day) but each bet carries genuine positive expected value. Quality over quantity.

ai-engine/reports/diagnostic_backtest_20260525_024437.csv

@@ -0,0 +1,51 @@
+match_id,match_date,league_id,score_home,score_away,ht_score_home,ht_score_away,market,pick,odds,stake_units,playable,won,unit_profit,raw_confidence,calibrated_confidence,play_score,ev_edge,bet_grade,is_value_sniper,bb_score,bb_action,bb_vetoes,bb_issues,bb_positives,bb_model_prob,bb_implied_prob,bb_model_market_gap,bb_divergence,bb_trap_market,v27_consensus,data_quality_score,data_quality_flags,risk_level,odds_reliability,htft_reversal_prob,htft_top_pick,league_name,is_cup,model_version,decision_reason
+5iam9c9dw3ggz3y1ohr9uh53o,2026-05-24,8nbwkj392b0xzssqpw9jwmzdn,0,0,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.33,missing_full_ms_odds;lineup_unavailable;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+8c90p2ft4zxjdck8wlgq1a61g,2026-05-24,a9vrdkelbgif0gtu3wxsr75xo,2,2,0,0,OU25,Üst,1.41,0.2,True,True,0.082,59.5,65.4,58.4,-0.1557,B,True,26.4,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.5955,0.7092,-0.1137,0.0339,True,DISAGREE,0.74,ai_features_inferred_from_history,MEDIUM,0.6618,,,,False,v28-pro-max,betting_brain_approved
+9ljz1grea3a8jajif4e9b7bpw,2026-05-24,2wolc27r8z03itcvwp43e38c5,1,1,1,1,BTTS,KG Var,1.62,0.2,True,True,0.124,53.7,69.9,48.4,-0.1476,B,True,1.6,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5371,0.6173,-0.0802,,True,AGREE,0.74,lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,HIGH,0.5592,,,,False,v28-pro-max,betting_brain_approved
+1r7iq0nhg2b674jpcm92ragpg,2026-05-24,1zp1du9n4rj36p1ss9zbxtqfb,4,0,1,0,BTTS,KG Var,1.89,0.2,True,False,-0.2,53.2,69.9,52.9,-0.0832,B,True,0.0,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5317,0.5291,0.0026,,True,AGREE,0.74,lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,HIGH,0.5353,,,,False,v28-pro-max,betting_brain_approved
+70dgok3yq76g076vemaps0178,2026-05-24,6lwpjhktjhl9g7x2w7njmzva6,2,1,1,0,BTTS,KG Var,1.74,0.2,True,True,0.148,51.9,68.4,60.4,-0.1255,B,True,15.7,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5192,0.5747,-0.0555,,True,AGREE,0.74,ai_features_inferred_from_history,MEDIUM,0.6164,,,,False,v28-pro-max,betting_brain_approved
+72q9d4uimmby6g6bor26taz9w,2026-05-24,6jgwiu2gq3dllmrwt45pfdn2z,2,0,2,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.555,missing_full_ms_odds;lineup_probable_not_confirmed;lineup_projection_low_confidence;lineup_incomplete;missing_referee;ai_features_inferred_from_history,LOW,,,,,False,v28-pro-max,no_bet_conditions_met
+67dd2t7043kv0yw1zj1buwdg4,2026-05-24,8n9w0n3i9kk05echhtmstn6o9,1,1,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.698,missing_full_ms_odds;lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+ejdwfph35q57phtfz7jr8st1w,2026-05-24,8ey0ww2zsosdmwr8ehsorh6t7,0,2,0,0,BTTS,KG Var,1.82,0.2,True,False,-0.2,53.2,69.9,63.5,-0.1097,B,True,12.1,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5317,0.5495,-0.0178,,True,DISAGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.51,,X/X,,False,v28-pro-max,betting_brain_approved
+68y8tlfnilw5trs1oqi4dhfkk,2026-05-24,8n9w0n3i9kk05echhtmstn6o9,2,2,1,1,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,DISAGREE,0.535,missing_full_ms_odds;lineup_probable_not_confirmed;lineup_projection_low_confidence;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+68fghiwdtwspk1m0ft5mspzx0,2026-05-24,8n9w0n3i9kk05echhtmstn6o9,1,2,1,2,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,DISAGREE,0.602,missing_full_ms_odds;lineup_probable_not_confirmed;lineup_projection_low_confidence;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+a1kqq0ggywfl6sl4srntxy2hg,2026-05-24,3ww12jab49q8q8mk9avdwjqgk,1,1,1,0,OU25,Üst,1.62,0.2,True,False,-0.2,61.4,65.4,60.6,-0.148,B,True,27.0,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.6144,0.6173,-0.0029,0.0854,True,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.5961,,,,False,v28-pro-max,betting_brain_approved
+a1vpp4i6t61v7qm3dfy6iuj2s,2026-05-24,3ww12jab49q8q8mk9avdwjqgk,2,1,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.74,missing_full_ms_odds;live_match_pre_match_features;ai_features_inferred_from_history,LOW,,,,,False,v28-pro-max,no_bet_conditions_met
+28rk26zqah60qj7h78qh5beac,2026-05-24,663a54fmymndjeev47qm7d3nf,2,2,2,1,OU25,Üst,1.39,0.2,True,True,0.078,61.1,65.4,57.5,-0.2039,B,True,13.3,BET,,low_reliability_league;inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.6107,0.7194,-0.1087,0.0577,True,AGREE,0.74,lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,MEDIUM,0.3522,,,,False,v28-pro-max,betting_brain_approved
+a26stt54g5ju5ecodpcbcqxw4,2026-05-24,3ww12jab49q8q8mk9avdwjqgk,2,2,1,1,BTTS,KG Var,2.05,0.2,True,True,0.21,53.7,69.9,70.6,0.0354,B,True,33.5,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;usable_historical_sample,0.5371,0.4878,0.0493,,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.5961,,,,False,v28-pro-max,betting_brain_approved
+318rrnl9b1vlenjd9hwahpvro,2026-05-24,9z5643nd06afqu01ea2wt8y4g,1,0,0,0,DC,1X,1.12,0.0,False,True,0.0,73.5,73.5,72.9,-0.0553,PASS,True,41.8,WATCH_NO_VALUE,odds_below_minimum,base_model_not_playable;inferred_statistical_features,value_sniper_override;v25_v27_aligned;triple_value_confirmed;usable_historical_sample,0.7351,0.8929,-0.1578,0.031,False,AGREE,0.51,lineup_unavailable;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,0.8734,,,,False,v28-pro-max,betting_brain_no_value_odds_below_minimum
+65tlfs3m6sc70261z37i90jys,2026-05-24,89ovpy1rarewwzqvi30bfdr8b,4,3,1,1,OU25,Üst,1.36,0.2,True,True,0.072,60.2,65.4,53.5,-0.2107,B,True,24.8,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.6018,0.7353,-0.1335,0.0982,True,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.7068,,,,False,v28-pro-max,betting_brain_approved
+ei6nilo5074tnb17kvv37oy6s,2026-05-24,3j81qr7yc4gdnakfwnxf95ovh,1,1,1,1,MS,2,1.11,0.0,False,False,0.0,75.5,90.9,66.9,-0.0497,PASS,True,0.0,WATCH_NO_VALUE,odds_below_minimum,base_model_not_playable;inferred_statistical_features;v25_v27_soft_disagreement;triple_value_not_confirmed,value_sniper_override;strong_historical_sample,0.7554,0.9009,-0.1455,0.204,False,AGREE,0.51,lineup_probable_not_confirmed;lineup_projection_low_confidence;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,0.8771,0.0,,,False,v28-pro-max,betting_brain_no_value_odds_below_minimum
+cx8nb7w3nmell38mta1umh2qc,2026-05-24,54c65mhi143utomzvvv3q2avh,0,2,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.726,missing_full_ms_odds;lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+53348iqniod61z7xurb4tx250,2026-05-24,477yyajzheg2z8u7uick0e13e,2,0,0,0,OU25,Üst,1.93,0.2,True,False,-0.2,59.5,65.4,71.3,0.0414,B,True,45.8,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.5955,0.5181,0.0774,0.1084,False,AGREE,0.74,lineup_probable_not_confirmed;ai_features_inferred_from_history,MEDIUM,0.4706,,,,False,v28-pro-max,betting_brain_approved
+2hw717s7fxi2v2w53kdsplhqs,2026-05-24,8najqkluatpaxvqws78b9s17c,1,1,0,0,DC,12,1.24,0.0,False,False,0.0,70.0,70.0,61.2,-0.12,PASS,True,0.0,WATCH_NO_VALUE,odds_below_minimum,base_model_not_playable;inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,value_sniper_override;v25_v27_aligned;strong_historical_sample,0.7001,0.8065,-0.1064,0.0075,True,AGREE,0.51,lineup_unavailable;lineup_incomplete;missing_referee;ai_features_inferred_from_history,HIGH,0.5082,,,,False,v28-pro-max,betting_brain_no_value_odds_below_minimum
+1q1s55dy4d4z4gs34qk6vx9n8,2026-05-24,cegl2ivkc25blcatxp4jmk1ec,1,0,1,0,OU25,Üst,2.03,0.2,True,False,-0.2,50.0,57.7,54.3,0.0546,B,True,35.7,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.5,0.4926,0.0074,0.0555,False,AGREE,0.74,lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,HIGH,0.5993,,,,False,v28-pro-max,betting_brain_approved
+648u8zd49cwcxpspmvinlmexg,2026-05-24,1eruend45vd20g9hbrpiggs5u,1,0,0,0,MS,1,1.32,0.2,True,True,0.064,63.1,65.5,55.7,-0.0828,B,True,39.4,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.6313,0.7576,-0.1263,0.0683,False,AGREE,0.74,lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,MEDIUM,0.8083,0.0,,,False,v28-pro-max,betting_brain_approved
+1psnufak57w8dfs9e5cvbmgwk,2026-05-24,cegl2ivkc25blcatxp4jmk1ec,3,0,2,0,BTTS,KG Var,1.94,0.2,True,False,-0.2,53.7,69.9,53.1,-0.0765,B,True,5.3,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5371,0.5155,0.0216,,True,AGREE,0.74,lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,HIGH,0.5993,,,,False,v28-pro-max,betting_brain_approved
+921kqaviappxt0w1kfmq1ek2c,2026-05-24,byu00jvt1j6csyv4y1lkt2fm2,1,0,1,0,DC,X2,1.18,0.0,False,False,0.0,75.8,75.8,74.9,-0.1267,PASS,True,0.0,WATCH_NO_VALUE,odds_below_minimum,inferred_statistical_features;v25_v27_soft_disagreement;trap_market_market_overpriced;triple_value_not_confirmed;engine_consensus_disagree,base_model_playable;value_sniper_override;strong_historical_sample,0.7584,0.8475,-0.0891,0.1783,True,DISAGREE,0.74,ai_features_inferred_from_history,MEDIUM,0.5359,,,,False,v28-pro-max,betting_brain_no_value_odds_below_minimum
+3m11hvh2fzailt3ykd0uhzz84,2026-05-24,54c65mhi143utomzvvv3q2avh,0,0,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.591,missing_full_ms_odds;lineup_probable_not_confirmed;lineup_projection_low_confidence;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+7kvvf6blnps2xk15100ccdedw,2026-05-24,4zwgbb66rif2spcoeeol2motx,5,0,3,0,BTTS,KG Var,1.33,0.2,True,False,-0.2,57.1,69.9,62.3,-0.1512,B,True,29.1,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5707,0.7519,-0.1812,,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.6995,,1/1,,False,v28-pro-max,betting_brain_approved
+7liir8zj32o7m2udr7cknb8d0,2026-05-24,4zwgbb66rif2spcoeeol2motx,3,0,2,0,OU25,Üst,1.33,0.2,True,True,0.066,61.4,65.4,58.3,-0.1437,B,True,27.0,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.6144,0.7519,-0.1375,0.0279,True,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.6995,,1/1,,False,v28-pro-max,betting_brain_approved
+7l74ilyz7olljclexvn8tbjtg,2026-05-24,4zwgbb66rif2spcoeeol2motx,5,1,4,0,BTTS,KG Var,1.55,0.2,True,True,0.11,57.1,69.9,68.2,-0.0873,B,True,32.7,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5707,0.6452,-0.0745,,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.6995,,1/1,,False,v28-pro-max,betting_brain_approved
+8f6gex4eh119d2hh9y2zb5clw,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,2,0,2,0,OU25,Üst,1.49,0.2,True,False,-0.2,66.5,65.4,77.9,0.0081,B,True,50.1,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.6651,0.6711,-0.006,0.1144,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,LOW,0.5033,,1/1,,False,v28-pro-max,betting_brain_approved
+8ee7ipt4u6kyk6baueedsdafo,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,0,2,0,2,BTTS,KG Var,1.69,0.2,True,False,-0.2,54.3,69.9,64.7,-0.0792,B,True,27.6,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5433,0.5917,-0.0484,,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.5033,,2/2,,False,v28-pro-max,betting_brain_approved
+8fydg367drpc25r1bobxqj3f8,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,3,1,2,1,OU25,Üst,1.61,0.2,True,True,0.122,50.1,57.7,50.7,-0.1072,B,True,35.0,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.5009,0.6211,-0.1202,0.0553,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.5033,,,,False,v28-pro-max,betting_brain_approved
+8fkdhce1peguwgnsunwoln3f8,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,0,2,0,2,OU25,Üst,1.24,0.0,False,False,0.0,61.4,65.4,61.1,-0.1548,PASS,True,16.8,WATCH_NO_VALUE,odds_below_minimum,inferred_statistical_features;triple_value_not_confirmed;historical_sample_too_low,base_model_playable;value_sniper_override;v25_v27_aligned,0.6144,0.8065,-0.1921,0.0942,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,LOW,0.5033,,,,False,v28-pro-max,betting_brain_no_value_odds_below_minimum
+9g5hqtjja6ceqhkpghwmoy6ms,2026-05-24,2y8bntiif3a9y6gtmauv30gt,2,0,1,0,OU25,Üst,1.71,0.2,True,False,-0.2,50.1,57.7,52.6,-0.0794,B,True,36.5,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.5009,0.5848,-0.0839,0.0481,False,DISAGREE,0.74,ai_features_inferred_from_history,MEDIUM,0.4782,,,,False,v28-pro-max,betting_brain_approved
+8h6429zr5ijqcxc8gjxygjtw4,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,3,0,1,0,MS,1,1.33,0.2,True,True,0.066,66.1,65.5,66.7,-0.1155,B,True,0.0,BET,,inferred_statistical_features;v25_v27_soft_disagreement;trap_market_market_overpriced;triple_value_not_confirmed;htft_reversal_prob_minor=0.11,base_model_playable;value_sniper_override;strong_historical_sample,0.6614,0.7519,-0.0905,0.2583,True,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.5033,0.0814,1/1,,False,v28-pro-max,betting_brain_approved
+77knm2ibdtb7akzrbltwz7axg,2026-05-24,bly7ema5au6j40i0grhl0pnub,1,1,1,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.726,missing_full_ms_odds;lineup_probable_not_confirmed;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+8es4680yd87gtmomg2jk3isyc,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,0,1,0,0,OU25,Üst,1.53,0.2,True,False,-0.2,59.5,65.4,67.8,-0.0259,B,True,42.4,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.5955,0.6536,-0.0581,0.0713,False,DISAGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.5033,,2/2,,False,v28-pro-max,betting_brain_approved
+8dmcz3k1u4ze53nvrsoz7eoes,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,1,1,1,0,BTTS,KG Var,1.26,0.0,False,True,0.0,54.9,69.9,57.6,-0.2005,PASS,True,3.3,WATCH_NO_VALUE,odds_below_minimum,inferred_statistical_features;triple_value_not_confirmed;historical_sample_too_low,base_model_playable;value_sniper_override,0.5488,0.7937,-0.2449,,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,LOW,0.5033,,1/1,,False,v28-pro-max,betting_brain_no_value_odds_below_minimum
+8gcbai6m7v7o8piqfram4qe50,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,3,1,2,0,HTFT,1/1,4.59,0.0,False,True,0.0,27.3,27.3,24.6,0.1657,PASS,True,0.0,REJECT,calibrated_confidence_too_low;play_score_too_low;volatile_market_requires_exceptional_evidence,inferred_statistical_features;historical_sample_too_low,base_model_playable,0.2734,0.2179,0.0555,,False,DISAGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.5033,,1/1,,False,v28-pro-max,betting_brain_no_safe_pick
+3azy3enp78au0zfugc3l1yf4k,2026-05-24,54c65mhi143utomzvvv3q2avh,2,0,1,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.532,missing_full_ms_odds;lineup_probable_not_confirmed;lineup_projection_low_confidence;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+1d2fb7bt5f8xy5on24w1kj1g4,2026-05-24,54c65mhi143utomzvvv3q2avh,1,0,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.532,missing_full_ms_odds;lineup_probable_not_confirmed;lineup_projection_low_confidence;lineup_incomplete;missing_referee;ai_features_inferred_from_history,LOW,,,,,False,v28-pro-max,no_bet_conditions_met
+pw01xm8v3jlz13fpi3zq0ftg,2026-05-24,3umprqta6ipyann6qjjh07biz,1,1,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.33,missing_full_ms_odds;lineup_unavailable;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+mjo9k4zr1x884vjlwea2y1hw,2026-05-24,3umprqta6ipyann6qjjh07biz,1,0,1,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.33,missing_full_ms_odds;lineup_unavailable;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+8d8fm7wli7tfx8hm9w5l8nuhg,2026-05-24,3is4bkgf3loxv9qfg3hm8zfqb,1,1,1,0,BTTS,KG Var,1.72,0.2,True,True,0.144,53.7,69.9,65.2,-0.0712,B,True,28.0,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5371,0.5814,-0.0443,,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.5033,,,,False,v28-pro-max,betting_brain_approved
+oqsq3f0kvic8xfed8dp302z8,2026-05-24,3umprqta6ipyann6qjjh07biz,3,2,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.33,missing_full_ms_odds;lineup_unavailable;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+o7tn4si7fxvq9c2mg0xs48wk,2026-05-24,3umprqta6ipyann6qjjh07biz,0,1,0,0,,,,1.0,False,,0.0,,,,,,False,,,,,,,,,,False,AGREE,0.33,missing_full_ms_odds;lineup_unavailable;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,,,,,False,v28-pro-max,no_bet_conditions_met
+eh9jfegscokidyczxfq691990,2026-05-24,3j81qr7yc4gdnakfwnxf95ovh,2,3,0,1,OU25,Üst,1.44,0.2,True,True,0.088,50.1,57.7,32.9,-0.2537,B,True,17.0,BET,,inferred_statistical_features;trap_market_market_overpriced;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.5009,0.6944,-0.1935,0.0596,True,AGREE,0.51,lineup_unavailable;lineup_incomplete;missing_referee;ai_features_inferred_from_history,MEDIUM,0.8771,,,,False,v28-pro-max,betting_brain_approved
+dkhhkbwnxwl47e8hybv89mwb8,2026-05-24,5jd0k2txwnq69frs79eulba8j,1,2,0,1,OU25,Üst,1.23,0.0,False,True,0.0,61.4,65.4,61.2,-0.1185,PASS,True,11.4,WATCH_NO_VALUE,odds_below_minimum,base_model_not_playable;inferred_statistical_features;triple_value_not_confirmed,value_sniper_override;v25_v27_aligned;strong_historical_sample,0.6144,0.813,-0.1986,0.0179,False,AGREE,0.74,ai_features_inferred_from_history,LOW,0.9233,,1/1,,False,v28-pro-max,betting_brain_no_value_odds_below_minimum
+1lknqdz9vmb3hnqu144zkkefo,2026-05-24,1r097lpxe0xn03ihb7wi98kao,1,0,1,0,BTTS,KG Var,1.78,0.2,True,False,-0.2,50.0,61.7,55.6,-0.088,B,True,29.3,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5,0.5618,-0.0618,,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,MEDIUM,0.7391,,1/1,,False,v28-pro-max,betting_brain_approved
+3oazp9kfbyyiatn246k4to6xg,2026-05-24,9ynnnx1qmkizq1o3qr3v0nsuk,1,2,0,1,BTTS,KG Var,1.36,0.2,True,True,0.072,53.7,69.9,61.2,-0.1571,B,True,33.7,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;strong_historical_sample,0.5371,0.7353,-0.1982,,False,AGREE,0.74,live_match_pre_match_features;ai_features_inferred_from_history,LOW,0.554,,2/2,,False,v28-pro-max,betting_brain_approved
+8cr8t6qh0r6g0mv6ftq0ic1sk,2026-05-24,a9vrdkelbgif0gtu3wxsr75xo,2,1,0,1,OU25,Üst,1.46,0.2,True,True,0.092,61.4,65.4,68.1,-0.0182,B,True,47.8,BET,,inferred_statistical_features;triple_value_not_confirmed,base_model_playable;value_sniper_override;v25_v27_aligned;strong_historical_sample,0.6144,0.6849,-0.0705,0.0535,False,AGREE,0.74,ai_features_inferred_from_history,MEDIUM,0.6618,,,,False,v28-pro-max,betting_brain_approved

ai-engine/reports/diagnostic_backtest_20260525_024437.json

@@ -0,0 +1,220 @@
+{
+  "args": {
+    "days": 3,
+    "max_matches": 50,
+    "start": null,
+    "end": null,
+    "progress_interval": 50
+  },
+  "aggregate": {
+    "overall": {
+      "n_total": 50,
+      "n_playable_settled": 27,
+      "wins": 15,
+      "losses": 12,
+      "hit_rate_pct": 55.56,
+      "unit_profit": -0.862,
+      "staked": 5.4,
+      "roi_pct": -15.96
+    },
+    "by_market": {
+      "OU25": {
+        "n_total": 13,
+        "n_playable_settled": 13,
+        "wins": 7,
+        "losses": 6,
+        "hit_rate_pct": 53.85,
+        "unit_profit": -0.6,
+        "staked": 2.6,
+        "roi_pct": -23.08
+      },
+      "BTTS": {
+        "n_total": 12,
+        "n_playable_settled": 12,
+        "wins": 6,
+        "losses": 6,
+        "hit_rate_pct": 50.0,
+        "unit_profit": -0.392,
+        "staked": 2.4,
+        "roi_pct": -16.33
+      },
+      "MS": {
+        "n_total": 2,
+        "n_playable_settled": 2,
+        "wins": 2,
+        "losses": 0,
+        "hit_rate_pct": 100.0,
+        "unit_profit": 0.13,
+        "staked": 0.4,
+        "roi_pct": 32.5
+      }
+    },
+    "by_confidence": {
+      "65-70": {
+        "n_total": 22,
+        "n_playable_settled": 22,
+        "wins": 13,
+        "losses": 9,
+        "hit_rate_pct": 59.09,
+        "unit_profit": -0.472,
+        "staked": 4.4,
+        "roi_pct": -10.73
+      },
+      "55-60": {
+        "n_total": 4,
+        "n_playable_settled": 4,
+        "wins": 2,
+        "losses": 2,
+        "hit_rate_pct": 50.0,
+        "unit_profit": -0.19,
+        "staked": 0.8,
+        "roi_pct": -23.75
+      },
+      "60-65": {
+        "n_total": 1,
+        "n_playable_settled": 1,
+        "wins": 0,
+        "losses": 1,
+        "hit_rate_pct": 0.0,
+        "unit_profit": -0.2,
+        "staked": 0.2,
+        "roi_pct": -100.0
+      }
+    },
+    "by_odds": {
+      "1.3-1.5": {
+        "n_total": 11,
+        "n_playable_settled": 11,
+        "wins": 9,
+        "losses": 2,
+        "hit_rate_pct": 81.82,
+        "unit_profit": 0.28,
+        "staked": 2.2,
+        "roi_pct": 12.73
+      },
+      "1.5-1.8": {
+        "n_total": 10,
+        "n_playable_settled": 10,
+        "wins": 5,
+        "losses": 5,
+        "hit_rate_pct": 50.0,
+        "unit_profit": -0.352,
+        "staked": 2.0,
+        "roi_pct": -17.6
+      },
+      "1.8-2.2": {
+        "n_total": 6,
+        "n_playable_settled": 6,
+        "wins": 1,
+        "losses": 5,
+        "hit_rate_pct": 16.67,
+        "unit_profit": -0.79,
+        "staked": 1.2,
+        "roi_pct": -65.83
+      }
+    },
+    "by_grade": {
+      "B": {
+        "n_total": 27,
+        "n_playable_settled": 27,
+        "wins": 15,
+        "losses": 12,
+        "hit_rate_pct": 55.56,
+        "unit_profit": -0.862,
+        "staked": 5.4,
+        "roi_pct": -15.96
+      }
+    },
+    "by_competition": {
+      "league": {
+        "n_total": 27,
+        "n_playable_settled": 27,
+        "wins": 15,
+        "losses": 12,
+        "hit_rate_pct": 55.56,
+        "unit_profit": -0.862,
+        "staked": 5.4,
+        "roi_pct": -15.96
+      }
+    }
+  },
+  "loss_diagnostics": {
+    "n_losses": 12,
+    "total_loss_units": -2.4,
+    "patterns": {
+      "high_htft_reversal_prob (>=0.20)": [
+        0,
+        0.0
+      ],
+      "cup_match": [
+        0,
+        0.0
+      ],
+      "low_league_reliability (<0.45)": [
+        0,
+        0.0
+      ],
+      "v27_disagree": [
+        3,
+        25.0
+      ],
+      "trap_market_flagged": [
+        4,
+        33.33
+      ],
+      "low_calibrated_conf (<55)": [
+        0,
+        0.0
+      ],
+      "high_odds_underdog (>=2.5)": [
+        0,
+        0.0
+      ],
+      "low_data_quality (<0.55)": [
+        0,
+        0.0
+      ],
+      "high_risk_level": [
+        3,
+        25.0
+      ],
+      "inferred_features": [
+        12,
+        100.0
+      ]
+    },
+    "by_market": [
+      [
+        "BTTS",
+        6
+      ],
+      [
+        "OU25",
+        6
+      ]
+    ],
+    "by_league": [
+      [
+        null,
+        12
+      ]
+    ],
+    "top_bb_issues_in_losses": [
+      [
+        "inferred_statistical_features",
+        12
+      ],
+      [
+        "triple_value_not_confirmed",
+        12
+      ],
+      [
+        "trap_market_market_overpriced",
+        4
+      ]
+    ],
+    "top_bb_vetoes_in_losses": []
+  },
+  "recommendations": [],
+  "errors_sample": []
+}

ai-engine/reports/diagnostic_backtest_20260525_024437.txt

@@ -0,0 +1,71 @@
+==============================================================================
+DIAGNOSTIC BACKTEST REPORT
+==============================================================================
+Generated:      2026-05-25T02:44:37
+Sample window:  start=-3d, end=now
+Max matches:    50
+Excluded days:  ['2026-04-29', '2026-05-03']
+
+OVERALL
+------------------------------------------------------------------------------
+  n_total                  : 50
+  n_playable_settled       : 27
+  wins                     : 15
+  losses                   : 12
+  hit_rate_pct             : 55.56
+  unit_profit              : -0.862
+  staked                   : 5.4
+  roi_pct                  : -15.96
+
+PER MARKET
+------------------------------------------------------------------------------
+  market        n    hit%    profit    roi%
+  OU25         13   53.85      -0.6  -23.08
+  BTTS         12    50.0    -0.392  -16.33
+  MS            2   100.0      0.13    32.5
+
+PER CALIBRATED CONFIDENCE BAND
+------------------------------------------------------------------------------
+  band            n    hit%    roi%
+  55-60           4    50.0  -23.75
+  60-65           1     0.0  -100.0
+  65-70          22   59.09  -10.73
+
+PER ODDS BAND
+------------------------------------------------------------------------------
+  band            n    hit%    roi%
+  1.3-1.5        11   81.82   12.73
+  1.5-1.8        10    50.0   -17.6
+  1.8-2.2         6   16.67  -65.83
+
+LEAGUE vs CUP
+------------------------------------------------------------------------------
+  league   n=  27 hit=55.56% roi=-15.96%
+
+LOSS DIAGNOSTICS
+------------------------------------------------------------------------------
+  total losses:        12
+  total lost units:    -2.4
+  By market:           [('BTTS', 6), ('OU25', 6)]
+  Loss patterns (count, % of losses):
+    high_htft_reversal_prob (>=0.20)                           0 (0.0%)
+    cup_match                                                  0 (0.0%)
+    low_league_reliability (<0.45)                             0 (0.0%)
+    v27_disagree                                               3 (25.0%)
+    trap_market_flagged                                        4 (33.33%)
+    low_calibrated_conf (<55)                                  0 (0.0%)
+    high_odds_underdog (>=2.5)                                 0 (0.0%)
+    low_data_quality (<0.55)                                   0 (0.0%)
+    high_risk_level                                            3 (25.0%)
+    inferred_features                                         12 (100.0%)
+  Top betting_brain issues seen in losses:
+    inferred_statistical_features                           12
+    triple_value_not_confirmed                              12
+    trap_market_market_overpriced                           4
+  Top betting_brain vetoes (in losses — i.e. veto fired but bet still went through value-sniper override):
+
+RECOMMENDATIONS
+------------------------------------------------------------------------------
+  (none surfaced — sample too small or no clear pattern)
+
+==============================================================================

ai-engine/reports/diagnostic_backtest_20260525_035649.json

@@ -0,0 +1,335 @@
+{
+  "args": {
+    "days": 14,
+    "max_matches": 1000,
+    "start": null,
+    "end": null,
+    "progress_interval": 50
+  },
+  "aggregate": {
+    "overall": {
+      "n_total": 1000,
+      "n_playable_settled": 524,
+      "wins": 287,
+      "losses": 237,
+      "hit_rate_pct": 54.77,
+      "unit_profit": -17.897,
+      "staked": 107.0,
+      "roi_pct": -16.73
+    },
+    "by_market": {
+      "OU25": {
+        "n_total": 236,
+        "n_playable_settled": 236,
+        "wins": 134,
+        "losses": 102,
+        "hit_rate_pct": 56.78,
+        "unit_profit": -6.271,
+        "staked": 48.5,
+        "roi_pct": -12.93
+      },
+      "BTTS": {
+        "n_total": 205,
+        "n_playable_settled": 205,
+        "wins": 105,
+        "losses": 100,
+        "hit_rate_pct": 51.22,
+        "unit_profit": -8.89,
+        "staked": 41.1,
+        "roi_pct": -21.63
+      },
+      "MS": {
+        "n_total": 76,
+        "n_playable_settled": 76,
+        "wins": 44,
+        "losses": 32,
+        "hit_rate_pct": 57.89,
+        "unit_profit": -2.396,
+        "staked": 16.0,
+        "roi_pct": -14.98
+      },
+      "OU35": {
+        "n_total": 3,
+        "n_playable_settled": 3,
+        "wins": 0,
+        "losses": 3,
+        "hit_rate_pct": 0.0,
+        "unit_profit": -0.6,
+        "staked": 0.6,
+        "roi_pct": -100.0
+      },
+      "DC": {
+        "n_total": 4,
+        "n_playable_settled": 4,
+        "wins": 4,
+        "losses": 0,
+        "hit_rate_pct": 100.0,
+        "unit_profit": 0.26,
+        "staked": 0.8,
+        "roi_pct": 32.5
+      }
+    },
+    "by_confidence": {
+      "65-70": {
+        "n_total": 420,
+        "n_playable_settled": 420,
+        "wins": 233,
+        "losses": 187,
+        "hit_rate_pct": 55.48,
+        "unit_profit": -14.057,
+        "staked": 85.1,
+        "roi_pct": -16.52
+      },
+      "60-65": {
+        "n_total": 33,
+        "n_playable_settled": 33,
+        "wins": 16,
+        "losses": 17,
+        "hit_rate_pct": 48.48,
+        "unit_profit": -1.61,
+        "staked": 6.6,
+        "roi_pct": -24.39
+      },
+      "55-60": {
+        "n_total": 52,
+        "n_playable_settled": 52,
+        "wins": 28,
+        "losses": 24,
+        "hit_rate_pct": 53.85,
+        "unit_profit": -0.668,
+        "staked": 10.5,
+        "roi_pct": -6.36
+      },
+      "50-55": {
+        "n_total": 5,
+        "n_playable_settled": 5,
+        "wins": 2,
+        "losses": 3,
+        "hit_rate_pct": 40.0,
+        "unit_profit": -0.64,
+        "staked": 1.3,
+        "roi_pct": -49.23
+      },
+      "45-50": {
+        "n_total": 8,
+        "n_playable_settled": 8,
+        "wins": 4,
+        "losses": 4,
+        "hit_rate_pct": 50.0,
+        "unit_profit": -0.382,
+        "staked": 1.9,
+        "roi_pct": -20.11
+      },
+      "70-80": {
+        "n_total": 6,
+        "n_playable_settled": 6,
+        "wins": 4,
+        "losses": 2,
+        "hit_rate_pct": 66.67,
+        "unit_profit": -0.54,
+        "staked": 1.6,
+        "roi_pct": -33.75
+      }
+    },
+    "by_odds": {
+      "1.3-1.5": {
+        "n_total": 241,
+        "n_playable_settled": 241,
+        "wins": 148,
+        "losses": 93,
+        "hit_rate_pct": 61.41,
+        "unit_profit": -7.408,
+        "staked": 49.0,
+        "roi_pct": -15.12
+      },
+      "1.5-1.8": {
+        "n_total": 221,
+        "n_playable_settled": 221,
+        "wins": 115,
+        "losses": 106,
+        "hit_rate_pct": 52.04,
+        "unit_profit": -6.926,
+        "staked": 44.3,
+        "roi_pct": -15.63
+      },
+      "1.8-2.2": {
+        "n_total": 56,
+        "n_playable_settled": 56,
+        "wins": 23,
+        "losses": 33,
+        "hit_rate_pct": 41.07,
+        "unit_profit": -2.789,
+        "staked": 12.2,
+        "roi_pct": -22.86
+      },
+      "2.2-3.0": {
+        "n_total": 5,
+        "n_playable_settled": 5,
+        "wins": 1,
+        "losses": 4,
+        "hit_rate_pct": 20.0,
+        "unit_profit": -0.574,
+        "staked": 1.3,
+        "roi_pct": -44.15
+      },
+      "3.0-5.0": {
+        "n_total": 1,
+        "n_playable_settled": 1,
+        "wins": 0,
+        "losses": 1,
+        "hit_rate_pct": 0.0,
+        "unit_profit": -0.2,
+        "staked": 0.2,
+        "roi_pct": -100.0
+      }
+    },
+    "by_grade": {
+      "B": {
+        "n_total": 518,
+        "n_playable_settled": 518,
+        "wins": 285,
+        "losses": 233,
+        "hit_rate_pct": 55.02,
+        "unit_profit": -16.931,
+        "staked": 105.3,
+        "roi_pct": -16.08
+      },
+      "A": {
+        "n_total": 6,
+        "n_playable_settled": 6,
+        "wins": 2,
+        "losses": 4,
+        "hit_rate_pct": 33.33,
+        "unit_profit": -0.966,
+        "staked": 1.7,
+        "roi_pct": -56.82
+      }
+    },
+    "by_competition": {
+      "league": {
+        "n_total": 524,
+        "n_playable_settled": 524,
+        "wins": 287,
+        "losses": 237,
+        "hit_rate_pct": 54.77,
+        "unit_profit": -17.897,
+        "staked": 107.0,
+        "roi_pct": -16.73
+      }
+    }
+  },
+  "loss_diagnostics": {
+    "n_losses": 237,
+    "total_loss_units": -48.7,
+    "patterns": {
+      "high_htft_reversal_prob (>=0.20)": [
+        0,
+        0.0
+      ],
+      "cup_match": [
+        0,
+        0.0
+      ],
+      "low_league_reliability (<0.45)": [
+        42,
+        17.72
+      ],
+      "v27_disagree": [
+        60,
+        25.32
+      ],
+      "trap_market_flagged": [
+        81,
+        34.18
+      ],
+      "low_calibrated_conf (<55)": [
+        7,
+        2.95
+      ],
+      "high_odds_underdog (>=2.5)": [
+        4,
+        1.69
+      ],
+      "low_data_quality (<0.55)": [
+        40,
+        16.88
+      ],
+      "high_risk_level": [
+        20,
+        8.44
+      ],
+      "inferred_features": [
+        0,
+        0.0
+      ]
+    },
+    "by_market": [
+      [
+        "OU25",
+        102
+      ],
+      [
+        "BTTS",
+        100
+      ],
+      [
+        "MS",
+        32
+      ],
+      [
+        "OU35",
+        3
+      ]
+    ],
+    "by_league": [
+      [
+        null,
+        237
+      ]
+    ],
+    "top_bb_issues_in_losses": [
+      [
+        "triple_value_not_confirmed",
+        230
+      ],
+      [
+        "trap_market_market_overpriced",
+        81
+      ],
+      [
+        "low_reliability_league",
+        40
+      ],
+      [
+        "v25_v27_soft_disagreement",
+        10
+      ],
+      [
+        "engine_consensus_disagree",
+        5
+      ],
+      [
+        "historical_sample_too_low",
+        3
+      ],
+      [
+        "very_low_reliability_league",
+        2
+      ],
+      [
+        "htft_reversal_prob_minor=0.13",
+        1
+      ]
+    ],
+    "top_bb_vetoes_in_losses": []
+  },
+  "recommendations": [
+    {
+      "type": "raise_confidence_threshold",
+      "confidence_band": "65-70",
+      "evidence": "n=420, roi=-16.52%",
+      "suggested_fix": "Raise MIN_BET_SCORE or market_min_conf above 65"
+    }
+  ],
+  "errors_sample": []
+}

ai-engine/reports/diagnostic_backtest_20260525_035649.txt

@@ -0,0 +1,86 @@
+==============================================================================
+DIAGNOSTIC BACKTEST REPORT
+==============================================================================
+Generated:      2026-05-25T03:56:49
+Sample window:  start=-14d, end=now
+Max matches:    1000
+Excluded days:  ['2026-04-29', '2026-05-03']
+
+OVERALL
+------------------------------------------------------------------------------
+  n_total                  : 1000
+  n_playable_settled       : 524
+  wins                     : 287
+  losses                   : 237
+  hit_rate_pct             : 54.77
+  unit_profit              : -17.897
+  staked                   : 107.0
+  roi_pct                  : -16.73
+
+PER MARKET
+------------------------------------------------------------------------------
+  market        n    hit%    profit    roi%
+  OU25        236   56.78    -6.271  -12.93
+  BTTS        205   51.22     -8.89  -21.63
+  MS           76   57.89    -2.396  -14.98
+  DC            4   100.0      0.26    32.5
+  OU35          3     0.0      -0.6  -100.0
+
+PER CALIBRATED CONFIDENCE BAND
+------------------------------------------------------------------------------
+  band            n    hit%    roi%
+  45-50           8    50.0  -20.11
+  50-55           5    40.0  -49.23
+  55-60          52   53.85   -6.36
+  60-65          33   48.48  -24.39
+  65-70         420   55.48  -16.52
+  70-80           6   66.67  -33.75
+
+PER ODDS BAND
+------------------------------------------------------------------------------
+  band            n    hit%    roi%
+  1.3-1.5       241   61.41  -15.12
+  1.5-1.8       221   52.04  -15.63
+  1.8-2.2        56   41.07  -22.86
+  2.2-3.0         5    20.0  -44.15
+  3.0-5.0         1     0.0  -100.0
+
+LEAGUE vs CUP
+------------------------------------------------------------------------------
+  league   n= 524 hit=54.77% roi=-16.73%
+
+LOSS DIAGNOSTICS
+------------------------------------------------------------------------------
+  total losses:        237
+  total lost units:    -48.7
+  By market:           [('OU25', 102), ('BTTS', 100), ('MS', 32), ('OU35', 3)]
+  Loss patterns (count, % of losses):
+    high_htft_reversal_prob (>=0.20)                           0 (0.0%)
+    cup_match                                                  0 (0.0%)
+    low_league_reliability (<0.45)                            42 (17.72%)
+    v27_disagree                                              60 (25.32%)
+    trap_market_flagged                                       81 (34.18%)
+    low_calibrated_conf (<55)                                  7 (2.95%)
+    high_odds_underdog (>=2.5)                                 4 (1.69%)
+    low_data_quality (<0.55)                                  40 (16.88%)
+    high_risk_level                                           20 (8.44%)
+    inferred_features                                          0 (0.0%)
+  Top betting_brain issues seen in losses:
+    triple_value_not_confirmed                              230
+    trap_market_market_overpriced                           81
+    low_reliability_league                                  40
+    v25_v27_soft_disagreement                               10
+    engine_consensus_disagree                               5
+    historical_sample_too_low                               3
+    very_low_reliability_league                             2
+    htft_reversal_prob_minor=0.13                           1
+  Top betting_brain vetoes (in losses — i.e. veto fired but bet still went through value-sniper override):
+
+RECOMMENDATIONS
+------------------------------------------------------------------------------
+  • [raise_confidence_threshold]
+      confidence_band: 65-70
+      evidence: n=420, roi=-16.52%
+      suggested_fix: Raise MIN_BET_SCORE or market_min_conf above 65
+
+==============================================================================

ai-engine/reports/filter_optimization_patch.json

@@ -0,0 +1,38 @@
+{
+  "BTTS": {
+    "min_calibrated_confidence": 65,
+    "min_ev_edge": -1.0,
+    "max_ev_edge": 0.1,
+    "min_odds": 1.4,
+    "max_odds": 10.0,
+    "min_odds_reliability": 0.55,
+    "require_v27_agree": true,
+    "expected_n_bets": 54,
+    "expected_hit_pct": 55.56,
+    "expected_roi_pct": -10.96
+  },
+  "MS": {
+    "min_calibrated_confidence": 0,
+    "min_ev_edge": -0.05,
+    "max_ev_edge": 0.15,
+    "min_odds": 1.2,
+    "max_odds": 10.0,
+    "min_odds_reliability": 0.0,
+    "require_v27_agree": true,
+    "expected_n_bets": 21,
+    "expected_hit_pct": 61.9,
+    "expected_roi_pct": 8.23
+  },
+  "OU25": {
+    "min_calibrated_confidence": 0,
+    "min_ev_edge": -1.0,
+    "max_ev_edge": 0.15,
+    "min_odds": 1.8,
+    "max_odds": 10.0,
+    "min_odds_reliability": 0.0,
+    "require_v27_agree": false,
+    "expected_n_bets": 20,
+    "expected_hit_pct": 65.0,
+    "expected_roi_pct": 28.91
+  }
+}

ai-engine/scripts/analyze_backtest_csv.py

@@ -0,0 +1,227 @@
+"""
+Deep root-cause analysis on diagnostic_backtest CSV.
+Tests specific hypotheses with hard numbers and proposes actionable
+filter rules with estimated impact (units saved, ROI shift).
+"""
+
+import sys, os, glob
+import pandas as pd
+import numpy as np
+
+REPORTS_DIR = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "reports")
+
+def latest_csv():
+    files = sorted(glob.glob(os.path.join(REPORTS_DIR, "diagnostic_backtest_*.csv")),
+                   key=os.path.getmtime, reverse=True)
+    return files[0] if files else None
+
+def fmt_pct(x):
+    return f"{x:>6.2f}%" if pd.notna(x) else "  ----"
+
+def cell(df, label, mask):
+    sub = df[mask]
+    n = len(sub)
+    if n == 0:
+        return f"  {label:<60} n=0"
+    wins = (sub["won"] == True).sum()
+    losses = (sub["won"] == False).sum()
+    settled = wins + losses
+    hr = 100.0 * wins / settled if settled else 0
+    profit = sub["unit_profit"].sum()
+    staked = sub["stake_units"].sum()
+    roi = 100.0 * profit / staked if staked else 0
+    return (f"  {label:<60} n={n:>4} hit={hr:>6.2f}% "
+            f"profit={profit:>+7.2f}u roi={roi:>+7.2f}%")
+
+def hypothesis_block(title, rows):
+    print(f"\n{'─' * 78}")
+    print(f"  {title}")
+    print(f"{'─' * 78}")
+    for row in rows:
+        print(row)
+
+def main():
+    csv_path = latest_csv()
+    if not csv_path:
+        print("No backtest CSV found")
+        return
+    print(f"Reading {csv_path}")
+    df = pd.read_csv(csv_path)
+    print(f"Loaded {len(df)} rows")
+
+    # Filter only playable + settled
+    pdf = df[(df["playable"] == True) & (df["won"].notna())].copy()
+    pdf["won"] = pdf["won"].astype(bool)
+    print(f"Playable + settled: {len(pdf)}")
+
+    overall_hr = (pdf["won"].sum() / len(pdf)) * 100
+    overall_roi = 100.0 * pdf["unit_profit"].sum() / pdf["stake_units"].sum()
+    print(f"\nOVERALL: hit={overall_hr:.2f}% roi={overall_roi:.2f}%")
+
+    # ─────────────────────────────────────────────────────────────────────
+    # H1: TRIPLE VALUE CONFIRMATION
+    # ─────────────────────────────────────────────────────────────────────
+    triple_confirmed_mask = ~pdf["bb_issues"].fillna("").str.contains(
+        "triple_value_not_confirmed", na=False
+    )
+    hypothesis_block(
+        "H1: TRIPLE VALUE CONFIRMED vs NOT CONFIRMED",
+        [
+            cell(pdf, "triple_value CONFIRMED", triple_confirmed_mask),
+            cell(pdf, "triple_value NOT CONFIRMED", ~triple_confirmed_mask),
+        ]
+    )
+
+    # ─────────────────────────────────────────────────────────────────────
+    # H2: TRAP MARKET FLAG
+    # ─────────────────────────────────────────────────────────────────────
+    trap_mask = pdf["bb_trap_market"] == True
+    hypothesis_block(
+        "H2: TRAP MARKET FLAG (model says band rate < implied → market overpriced)",
+        [
+            cell(pdf, "trap_market_flag = TRUE  (model warned)", trap_mask),
+            cell(pdf, "trap_market_flag = FALSE", ~trap_mask),
+        ]
+    )
+
+    # ─────────────────────────────────────────────────────────────────────
+    # H3: V25/V27 CONSENSUS
+    # ─────────────────────────────────────────────────────────────────────
+    agree_mask = pdf["v27_consensus"] == "AGREE"
+    disagree_mask = pdf["v27_consensus"] == "DISAGREE"
+    hypothesis_block(
+        "H3: V25 ↔ V27 CONSENSUS",
+        [
+            cell(pdf, "AGREE", agree_mask),
+            cell(pdf, "DISAGREE", disagree_mask),
+            cell(pdf, "neither/null", ~(agree_mask | disagree_mask)),
+        ]
+    )
+
+    # ─────────────────────────────────────────────────────────────────────
+    # H4: ODDS RELIABILITY (league quality)
+    # ─────────────────────────────────────────────────────────────────────
+    pdf["rel_band"] = pd.cut(
+        pdf["odds_reliability"].fillna(0.35),
+        [0, 0.30, 0.45, 0.55, 1.0],
+        labels=["<0.30 verylow", "0.30-0.45 low", "0.45-0.55 mid", ">=0.55 high"]
+    )
+    hypothesis_block(
+        "H4: LEAGUE ODDS RELIABILITY",
+        [cell(pdf, str(b), pdf["rel_band"] == b) for b in pdf["rel_band"].cat.categories]
+    )
+
+    # ─────────────────────────────────────────────────────────────────────
+    # H5: CALIBRATOR IMPACT (raw vs calibrated)
+    # ─────────────────────────────────────────────────────────────────────
+    pdf["calib_delta"] = pdf["calibrated_confidence"] - pdf["raw_confidence"]
+    pdf["delta_band"] = pd.cut(
+        pdf["calib_delta"].fillna(0),
+        [-100, -10, -3, 3, 10, 100],
+        labels=["cal<<raw (-10+)", "cal<raw (-3..-10)", "≈equal (±3)",
+                "cal>raw (3..10)", "cal>>raw (+10+)"]
+    )
+    hypothesis_block(
+        "H5: CALIBRATOR DELTA (calibrated_conf - raw_conf)",
+        [cell(pdf, str(b), pdf["delta_band"] == b) for b in pdf["delta_band"].cat.categories]
+    )
+
+    # ─────────────────────────────────────────────────────────────────────
+    # H6: EV EDGE
+    # ─────────────────────────────────────────────────────────────────────
+    pdf["edge_band"] = pd.cut(
+        pdf["ev_edge"].fillna(0),
+        [-10, -0.05, 0.0, 0.05, 0.10, 0.20, 10],
+        labels=["edge<-5%", "-5%-0%", "0-5%", "5-10%", "10-20%", ">20%"]
+    )
+    hypothesis_block(
+        "H6: EV EDGE (model_prob - implied_prob)",
+        [cell(pdf, str(b), pdf["edge_band"] == b) for b in pdf["edge_band"].cat.categories]
+    )
+
+    # ─────────────────────────────────────────────────────────────────────
+    # H7: ODDS x MARKET cross
+    # ─────────────────────────────────────────────────────────────────────
+    pdf["odds_band"] = pd.cut(
+        pdf["odds"].fillna(0),
+        [0, 1.30, 1.50, 1.80, 2.20, 3.00, 100],
+        labels=["<1.30", "1.30-1.50", "1.50-1.80", "1.80-2.20", "2.20-3.00", ">3.00"]
+    )
+
+    print(f"\n{'─' * 78}")
+    print(f"  H7: ODDS BAND × MARKET (per cell hit% / roi% / n)")
+    print(f"{'─' * 78}")
+    pivot_n = pdf.pivot_table(index="market", columns="odds_band",
+                                values="match_id", aggfunc="count", fill_value=0,
+                                observed=False)
+    pivot_roi = pdf.pivot_table(index="market", columns="odds_band",
+                                  values="unit_profit", aggfunc="sum", fill_value=0,
+                                  observed=False)
+    pivot_stake = pdf.pivot_table(index="market", columns="odds_band",
+                                    values="stake_units", aggfunc="sum", fill_value=0,
+                                    observed=False)
+    pivot_roi_pct = (100.0 * pivot_roi / pivot_stake.replace(0, np.nan)).round(1)
+    print("\n  Bet count per cell:")
+    print(pivot_n.to_string())
+    print("\n  ROI% per cell:")
+    print(pivot_roi_pct.to_string())
+
+    # ─────────────────────────────────────────────────────────────────────
+    # H8: COMBINED FILTER SIMULATION
+    # ─────────────────────────────────────────────────────────────────────
+    print(f"\n{'─' * 78}")
+    print("  H8: COMBINED FILTER SIMULATION (what if we add rules)")
+    print(f"{'─' * 78}")
+
+    def simulate(filter_name, keep_mask):
+        kept = pdf[keep_mask]
+        rejected = pdf[~keep_mask]
+        if len(kept) == 0:
+            return f"  {filter_name:<55} → 0 bet remain"
+        kept_hr = 100.0 * kept["won"].sum() / len(kept)
+        kept_profit = kept["unit_profit"].sum()
+        kept_staked = kept["stake_units"].sum()
+        kept_roi = 100.0 * kept_profit / kept_staked if kept_staked else 0
+        saved = -rejected["unit_profit"].sum()  # money we WOULD HAVE LOST
+        return (f"  {filter_name:<55} keep={len(kept):>3} hit={kept_hr:>5.1f}% "
+                f"roi={kept_roi:>+6.2f}% saved={saved:>+6.2f}u")
+
+    print(simulate("BASELINE (no extra filter)", pd.Series([True] * len(pdf), index=pdf.index)))
+    print(simulate("REJECT triple_value_not_confirmed",
+                   ~pdf["bb_issues"].fillna("").str.contains("triple_value_not_confirmed")))
+    print(simulate("REJECT trap_market_flag",
+                   ~(pdf["bb_trap_market"] == True)))
+    print(simulate("REJECT v27 DISAGREE",
+                   pdf["v27_consensus"] != "DISAGREE"))
+    print(simulate("REJECT odds_reliability < 0.45",
+                   pdf["odds_reliability"].fillna(1.0) >= 0.45))
+    print(simulate("REJECT odds in 1.80-2.20",
+                   (pdf["odds"].fillna(0) < 1.80) | (pdf["odds"].fillna(0) >= 2.20)))
+    print(simulate("REJECT ev_edge < 0",
+                   pdf["ev_edge"].fillna(0) >= 0))
+    print(simulate("REJECT ev_edge < 0.05",
+                   pdf["ev_edge"].fillna(0) >= 0.05))
+    print()
+    print("  COMBINED rules:")
+    # Stack 1: drop triple_not_confirmed + trap_market + DISAGREE
+    s1 = (
+        ~pdf["bb_issues"].fillna("").str.contains("triple_value_not_confirmed")
+        & ~(pdf["bb_trap_market"] == True)
+        & (pdf["v27_consensus"] != "DISAGREE")
+    )
+    print(simulate("STACK1: !triple_not_conf & !trap & !disagree", s1))
+    # Stack 2: + edge>=0
+    s2 = s1 & (pdf["ev_edge"].fillna(0) >= 0)
+    print(simulate("STACK2: STACK1 + edge >= 0", s2))
+    # Stack 3: + reliability>=0.45
+    s3 = s2 & (pdf["odds_reliability"].fillna(1.0) >= 0.45)
+    print(simulate("STACK3: STACK2 + reliability >= 0.45", s3))
+    # Stack 4: + odds outside 1.80-2.20
+    s4 = s3 & ((pdf["odds"].fillna(0) < 1.80) | (pdf["odds"].fillna(0) >= 2.20))
+    print(simulate("STACK4: STACK3 + odds NOT in 1.80-2.20", s4))
+
+    print(f"\n{'─' * 78}")
+    print("DONE.")
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/analyze_match_v2.py

@@ -0,0 +1,137 @@
+"""
+Analyze Match v2 — the per-match multi-market value board + disciplined pick.
+===========================================================================
+Answers "for ONE match, show every bet type's probability + model signal +
+market-vs-model value, and pick the right bet." Leak-free models.
+
+KEY HONEST RULE (proven by multi_market_edge.py): compute & SHOW value for all
+markets, but only MS (1X2) carries real, fold-consistent model edge. In OU/HT/
+BTTS the market is efficient — a big model-vs-market gap there is the MODEL'S
+ERROR, not value. So non-MS rows are INFO-ONLY; only an MS value bet in the
+favourite band is STAKED.
+
+Demo: trains all market models on the first 85% of history, then prints the full
+board for sample matches in the unseen last 15% (with what actually happened).
+
+Usage:
+  python scripts/analyze_match_v2.py --n 6
+  python scripts/analyze_match_v2.py --match <match_id>
+"""
+from __future__ import annotations
+import argparse, os, sys
+import numpy as np, pandas as pd, xgboost as xgb
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+META = {"match_id","home_team_id","away_team_id","league_id","mst_utc",
+        "score_home","score_away","ht_score_home","ht_score_away"}
+LEAKY = {"home_goals_form","away_goals_form","total_goals","ht_total_goals",
+         "squad_diff","home_squad_quality","away_squad_quality",
+         "referee_home_bias","referee_avg_goals"}
+STAKE_LO, STAKE_HI = 1.5, 2.4   # MS favourite band that staking is allowed in
+STAKE_MARGIN = 0.03
+
+def ou(line):   return lambda sh,sa,hh,ha: (0 if (sh+sa) > line else 1)
+def htou(line): return lambda sh,sa,hh,ha: (None if np.isnan(hh) else (0 if (hh+ha) > line else 1))
+MARKETS = {
+  "MS":      ("multi", ["odds_ms_h","odds_ms_d","odds_ms_a"], ["1","X","2"],
+              lambda sh,sa,hh,ha: 0 if sh>sa else (1 if sh==sa else 2)),
+  "OU25":    ("binary",["odds_ou25_o","odds_ou25_u"], ["2.5Üst","2.5Alt"], ou(2.5)),
+  "OU15":    ("binary",["odds_ou15_o","odds_ou15_u"], ["1.5Üst","1.5Alt"], ou(1.5)),
+  "OU35":    ("binary",["odds_ou35_o","odds_ou35_u"], ["3.5Üst","3.5Alt"], ou(3.5)),
+  "BTTS":    ("binary",["odds_btts_y","odds_btts_n"], ["KG Var","KG Yok"],
+              lambda sh,sa,hh,ha: 0 if (sh>0 and sa>0) else 1),
+  "HT":      ("multi", ["odds_ht_ms_h","odds_ht_ms_d","odds_ht_ms_a"], ["İY1","İYX","İY2"],
+              lambda sh,sa,hh,ha: None if np.isnan(hh) else (0 if hh>ha else (1 if hh==ha else 2))),
+  "HT_OU15": ("binary",["odds_ht_ou15_o","odds_ht_ou15_u"], ["İY1.5Üst","İY1.5Alt"], htou(1.5)),
+}
+STAKED_MARKETS = {"MS"}   # only these are bet; rest are info-only
+PM = {"objective":"multi:softprob","num_class":3,"max_depth":5,"eta":0.05,"subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+PB = {"objective":"binary:logistic","max_depth":5,"eta":0.05,"subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--n", type=int, default=6, help="how many sample matches")
+    ap.add_argument("--match", help="specific match_id")
+    ap.add_argument("--estimators", type=int, default=250)
+    args = ap.parse_args()
+
+    df = pd.read_csv(CSV, low_memory=False).sort_values("mst_utc").reset_index(drop=True)
+    sh = pd.to_numeric(df["score_home"],errors="coerce"); sa = pd.to_numeric(df["score_away"],errors="coerce")
+    ok = sh.notna()&sa.notna(); df = df[ok].reset_index(drop=True)
+    SH=sh[ok.values].values.astype(float); SA=sa[ok.values].values.astype(float)
+    HH=pd.to_numeric(df["ht_score_home"],errors="coerce").values.astype(float)
+    HA=pd.to_numeric(df["ht_score_away"],errors="coerce").values.astype(float)
+    feats=[c for c in df.columns if c not in META and not c.startswith("label_") and c not in LEAKY]
+    X=df[feats].apply(pd.to_numeric,errors="coerce").fillna(0.0).values
+    N=len(df); cut=int(N*0.85)
+    print(f"Training {len(MARKETS)} leak-free market models on {cut:,} matches ...")
+
+    models={}
+    for m,(kind,ocols,picks,tfn) in MARKETS.items():
+        if not all(c in df.columns for c in ocols): continue
+        truth=np.array([tfn(SH[i],SA[i],HH[i],HA[i]) for i in range(cut)],dtype=object)
+        valid=np.array([v is not None for v in truth])
+        if kind=="multi":
+            b=xgb.train(PM,xgb.DMatrix(X[:cut][valid],label=truth[valid].astype(int)),num_boost_round=args.estimators)
+        else:
+            b=xgb.train(PB,xgb.DMatrix(X[:cut][valid],label=(truth[valid].astype(int)==0).astype(int)),num_boost_round=args.estimators)
+        models[m]=(kind,ocols,picks,tfn,b)
+
+    # choose matches from holdout
+    hold = df.iloc[cut:].reset_index(drop=True)
+    if args.match:
+        sel_idx = df.index[df["match_id"].astype(str)==str(args.match)].tolist()
+        rows = [(i,) for i in sel_idx]
+        base = df
+    else:
+        pick_pos = np.linspace(0, len(hold)-1, args.n, dtype=int)
+        rows = [(cut+p,) for p in pick_pos]
+        base = df
+
+    for (gi,) in rows:
+        r = base.iloc[gi]
+        xrow = X[gi:gi+1]
+        sh_,sa_,hh_,ha_ = SH[gi],SA[gi],HH[gi],HA[gi]
+        ht = f"{int(hh_)}-{int(ha_)}" if not np.isnan(hh_) else "?"
+        print("\n"+"="*72)
+        print(f"MATCH {r['match_id']}  | elo H{r.get('home_overall_elo','?'):.0f} vs A{r.get('away_overall_elo','?'):.0f}"
+              f"  | ACTUAL {int(sh_)}-{int(sa_)} (HT {ht})")
+        print(f"  {'market':<8}{'pick':<10}{'model%':>8}{'impl%':>7}{'edge':>7}{'odds':>7}  flag   result")
+        print("  "+"-"*64)
+        best_ms=None
+        for m,(kind,ocols,picks,tfn,b) in models.items():
+            if kind=="multi":
+                P=b.predict(xgb.DMatrix(xrow))[0]
+            else:
+                p=float(b.predict(xgb.DMatrix(xrow))[0]); P=np.array([p,1-p])
+            O=pd.to_numeric(r[ocols],errors="coerce").fillna(0.0).values
+            truth=tfn(sh_,sa_,hh_,ha_)
+            for k in range(len(picks)):
+                o=O[k]
+                if o<=1.0: continue
+                imp=1.0/o; edge=P[k]-imp
+                res = "—" if truth is None else ("WON" if truth==k else "lost")
+                staked = (m in STAKED_MARKETS) and edge>STAKE_MARGIN and STAKE_LO<=o<STAKE_HI
+                flag = "★BET" if staked else ("val" if edge>STAKE_MARGIN else "")
+                print(f"  {m:<8}{picks[k]:<10}{100*P[k]:>7.1f}{100*imp:>7.1f}{100*edge:>+7.1f}{o:>7.2f}  {flag:<5} {res}")
+                if staked and (best_ms is None or edge>best_ms[0]):
+                    best_ms=(edge,m,picks[k],o,res)
+        print("  "+"-"*64)
+        if best_ms:
+            e,m,p,o,res = best_ms
+            print(f"  >>> STAKE: {m} {p} @ {o:.2f}  (edge +{100*e:.1f}%, favourite band)  -> {res}")
+        else:
+            print(f"  >>> NO STAKE: no MS value in favourite band. (Other markets info-only —")
+            print(f"      their 'value' is model error in efficient markets; do NOT chase it.)")
+    print("\nNOTE: only MS staked (proven edge). All markets shown for transparency.")
+    print("Forward-validate with CLV before real money. Static CSV odds may overstate edge.")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/betting_policy.py

@@ -0,0 +1,113 @@
+"""
+Betting Policy — the honest, leak-free strategy the data actually supports.
+==========================================================================
+Everything else in this repo bet UNDERDOGS (odds 6-7.5) and lost (-43.7% live).
+The data says the opposite: the only positive, fold-consistent, model-driven
+signal is MILD FAVOURITES the model rates above the market price.
+
+POLICY (MS / 1X2 only):
+  * leak-free model (drops the result-encoding features, see LEAKY)
+  * bet the model's single biggest value edge (model_prob - implied) ...
+  * ONLY if the picked side's odds are in [--lo, --hi]  (favourite band)
+  * ONLY if that edge > --margin
+  * flat 1u stake, one bet per match, never a longshot, never a parlay.
+
+Walk-forward, no leakage. Reports the policy ROI, fold consistency, drawdown,
+and the model-free baseline (blind favourite) so you can see the model's lift.
+
+⚠️ HONEST CAVEAT: CSV odds are a static capture, not the verified obtainable
+closing line. A small backtest edge here is a LEAD, not a guarantee. Forward
+paper-trade with real CLV (capture_closing_odds.py) before risking money.
+
+Usage: python scripts/betting_policy.py --lo 1.5 --hi 2.2 --margin 0.0 --folds 8
+"""
+from __future__ import annotations
+import argparse, os, sys
+import numpy as np, pandas as pd, xgboost as xgb
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+META = {"match_id","home_team_id","away_team_id","league_id","mst_utc",
+        "score_home","score_away","ht_score_home","ht_score_away"}
+LEAKY = {"home_goals_form","away_goals_form","total_goals","ht_total_goals",
+         "squad_diff","home_squad_quality","away_squad_quality",
+         "referee_home_bias","referee_avg_goals"}
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--lo", type=float, default=1.5)
+    ap.add_argument("--hi", type=float, default=2.2)
+    ap.add_argument("--margin", type=float, default=0.0)
+    ap.add_argument("--folds", type=int, default=8)
+    ap.add_argument("--estimators", type=int, default=250)
+    args = ap.parse_args()
+
+    df = pd.read_csv(CSV, low_memory=False).sort_values("mst_utc").reset_index(drop=True)
+    sh = pd.to_numeric(df["score_home"], errors="coerce")
+    sa = pd.to_numeric(df["score_away"], errors="coerce")
+    ok = sh.notna() & sa.notna()
+    df, sh, sa = df[ok].reset_index(drop=True), sh[ok.values].values, sa[ok.values].values
+    y = np.where(sh > sa, 0, np.where(sh == sa, 1, 2))
+    O = df[["odds_ms_h","odds_ms_d","odds_ms_a"]].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    feats = [c for c in df.columns if c not in META and not c.startswith("label_") and c not in LEAKY]
+    X = df[feats].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+
+    n = len(df); start = int(n*0.5)
+    bounds = np.linspace(start, n, args.folds+1, dtype=int)
+    params = {"objective":"multi:softprob","num_class":3,"max_depth":5,"eta":0.05,
+              "subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+
+    print(f"POLICY: favourite band [{args.lo},{args.hi}]  margin {args.margin}  "
+          f"leak-free feats={len(feats)}  folds={args.folds}\n")
+    all_pnl=[]; fold_rows=[]; base_pnl=[]
+    for fi in range(args.folds):
+        te0,te1 = bounds[fi], bounds[fi+1]
+        if te1-te0 < 50: continue
+        bst = xgb.train(params, xgb.DMatrix(X[:te0], label=y[:te0]), num_boost_round=args.estimators)
+        P = bst.predict(xgb.DMatrix(X[te0:te1]))
+        yte, Ote = y[te0:te1], O[te0:te1]
+        implied = np.where(Ote>1.0, 1.0/Ote, np.nan)
+        edge = np.where(np.isnan(implied), -9.0, P-implied)
+        pick = edge.argmax(1); pe = edge[np.arange(len(yte)),pick]; po = Ote[np.arange(len(yte)),pick]
+        bet = (pe>args.margin) & (po>=args.lo) & (po<args.hi)
+        win = (pick==yte)&bet
+        pnl = np.where(win, po-1.0, -1.0)[bet]
+        # model-free baseline: blind favourite in same band
+        fav=Ote.argmin(1); fo=Ote[np.arange(len(yte)),fav]
+        bmask=(fo>=args.lo)&(fo<args.hi)&(Ote>1.0).all(1)
+        bpnl=np.where(fav[bmask]==yte[bmask], fo[bmask]-1.0, -1.0)
+        roi = 100*pnl.sum()/len(pnl) if len(pnl) else float('nan')
+        broi= 100*bpnl.sum()/len(bpnl) if len(bpnl) else float('nan')
+        fold_rows.append((fi, len(pnl), 100*win.sum()/max(bet.sum(),1), roi, broi))
+        all_pnl.extend(pnl.tolist()); base_pnl.extend(bpnl.tolist())
+        print(f"  fold {fi}: policy_bets={len(pnl):>4}  hit={100*win.sum()/max(bet.sum(),1):>5.1f}%  "
+              f"ROI={roi:>7.2f}%   | baseline(blind fav) ROI={broi:>7.2f}%")
+
+    a=np.array(all_pnl); b=np.array(base_pnl)
+    print("\n"+"="*70)
+    print("AGGREGATE")
+    print("="*70)
+    if len(a):
+        cum=np.cumsum(a); peak=np.maximum.accumulate(cum); dd=(cum-peak).min()
+        folds_pos=sum(1 for r in fold_rows if r[3]>0)
+        print(f"  POLICY:   bets={len(a):>5}  hit={100*(a>0).mean():.1f}%  "
+              f"ROI={100*a.mean():+.2f}%  net={a.sum():+.1f}u  maxDD={dd:.1f}u  "
+              f"folds+={folds_pos}/{len(fold_rows)}")
+    if len(b):
+        print(f"  BASELINE: bets={len(b):>5}  hit={100*(b>0).mean():.1f}%  "
+              f"ROI={100*b.mean():+.2f}%  (blind favourite, same band)")
+    if len(a):
+        print(f"\n  MODEL LIFT over blind favourite: "
+              f"{100*a.mean()-100*b.mean():+.1f} percentage points")
+    print("\nREAD: a believable system has ROI>0, folds+ near full, tolerable maxDD,")
+    print("and clearly beats the blind-favourite baseline. Even then it's a LEAD —")
+    print("forward paper-trade with real CLV before staking real money.")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/calibration_report.py

@@ -0,0 +1,112 @@
+"""
+Calibration Report — are the model's probabilities "kusursuz"?
+=============================================================
+"Flawless probability" has a precise technical meaning: CALIBRATION. When the
+model says 60%, the event must happen ~60% of the time. This measures exactly
+that for the leak-free MS (1X2) model, and shows how much isotonic calibration
+improves it.
+
+Metrics:
+  * Reliability table: bin predicted prob -> avg predicted vs ACTUAL frequency.
+    Calibrated = avg_pred ≈ actual in every bin (gap ≈ 0).
+  * ECE  (Expected Calibration Error): weighted mean |pred - actual|. Lower=better.
+  * Brier score, Log-loss: overall probability accuracy. Lower=better.
+
+Time-split (no leakage): train 70% -> fit isotonic on next 15% -> test last 15%.
+
+Usage: python scripts/calibration_report.py
+"""
+from __future__ import annotations
+import os, sys
+import numpy as np, pandas as pd, xgboost as xgb
+from sklearn.isotonic import IsotonicRegression
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+META = {"match_id","home_team_id","away_team_id","league_id","mst_utc",
+        "score_home","score_away","ht_score_home","ht_score_away"}
+LEAKY = {"home_goals_form","away_goals_form","total_goals","ht_total_goals",
+         "squad_diff","home_squad_quality","away_squad_quality",
+         "referee_home_bias","referee_avg_goals"}
+PARAMS = {"objective":"multi:softprob","num_class":3,"max_depth":5,"eta":0.05,
+          "subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+
+
+def reliability(probs, y, nbins=10):
+    """Pool one-vs-rest predictions; bin by predicted prob; compare to actual freq."""
+    P = probs.reshape(-1)
+    hit = np.zeros((len(y), probs.shape[1]))
+    hit[np.arange(len(y)), y] = 1.0
+    H = hit.reshape(-1)
+    edges = np.linspace(0, 1, nbins + 1)
+    rows, ece, N = [], 0.0, len(P)
+    for i in range(nbins):
+        lo, hi = edges[i], edges[i+1]
+        m = (P >= lo) & (P < hi) if i < nbins-1 else (P >= lo) & (P <= hi)
+        if m.sum() == 0:
+            continue
+        ap, af, n = P[m].mean(), H[m].mean(), int(m.sum())
+        rows.append((f"{int(lo*100)}-{int(hi*100)}%", n, ap, af, af-ap))
+        ece += (n / N) * abs(ap - af)
+    return rows, ece
+
+
+def brier(probs, y):
+    oh = np.zeros_like(probs); oh[np.arange(len(y)), y] = 1.0
+    return float(np.mean(np.sum((probs - oh) ** 2, axis=1)))
+
+
+def logloss(probs, y):
+    p = np.clip(probs[np.arange(len(y)), y], 1e-9, 1)
+    return float(-np.mean(np.log(p)))
+
+
+def main():
+    df = pd.read_csv(CSV, low_memory=False).sort_values("mst_utc").reset_index(drop=True)
+    sh = pd.to_numeric(df["score_home"], errors="coerce")
+    sa = pd.to_numeric(df["score_away"], errors="coerce")
+    ok = sh.notna() & sa.notna()
+    df, sh, sa = df[ok].reset_index(drop=True), sh[ok.values].values, sa[ok.values].values
+    y = np.where(sh > sa, 0, np.where(sh == sa, 1, 2))
+    feats = [c for c in df.columns if c not in META and not c.startswith("label_") and c not in LEAKY]
+    X = df[feats].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+
+    n = len(df); a, b = int(n*0.70), int(n*0.85)
+    Xtr, ytr = X[:a], y[:a]
+    Xca, yca = X[a:b], y[a:b]
+    Xte, yte = X[b:], y[b:]
+    print(f"{n:,} matches | train {len(ytr):,} / calib {len(yca):,} / test {len(yte):,} (time-split)")
+
+    bst = xgb.train(PARAMS, xgb.DMatrix(Xtr, label=ytr), num_boost_round=300)
+    raw_ca = bst.predict(xgb.DMatrix(Xca))
+    raw_te = bst.predict(xgb.DMatrix(Xte))
+
+    # isotonic per class (fit on calib), apply to test, renormalize
+    isos = []
+    for k in range(3):
+        ir = IsotonicRegression(out_of_bounds="clip", y_min=0, y_max=1)
+        ir.fit(raw_ca[:, k], (yca == k).astype(float))
+        isos.append(ir)
+    cal_te = np.column_stack([isos[k].predict(raw_te[:, k]) for k in range(3)])
+    cal_te = np.clip(cal_te, 1e-6, 1)
+    cal_te = cal_te / cal_te.sum(axis=1, keepdims=True)
+
+    for name, P in (("RAW (kalibrasyonsuz)", raw_te), ("ISOTONIC KALİBRELİ", cal_te)):
+        rows, ece = reliability(P, yte)
+        print(f"\n{'='*64}\n{name}\n{'='*64}")
+        print(f"  {'tahmin bandı':<12}{'n':>7}{'ort.tahmin':>12}{'gerçek':>9}{'fark':>8}")
+        for band, nn, ap, af, gap in rows:
+            print(f"  {band:<12}{nn:>7}{100*ap:>11.1f}%{100*af:>8.1f}%{100*gap:>+7.1f}")
+        print(f"  ECE={100*ece:.2f}%   Brier={brier(P,yte):.4f}   LogLoss={logloss(P,yte):.4f}")
+
+    print("\nOKUMA: 'fark' ≈ 0 ise olasılıklar KUSURSUZ (söylediği %X gerçekten %X).")
+    print("ECE/Brier/LogLoss düştüyse kalibrasyon işe yaradı. Bu kalibre olasılıklar,")
+    print("maçın olası sonuçlarını dürüstçe gösterir — kayıp-minimizasyonun temeli budur.")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/calibration_scoreboard.py

@@ -0,0 +1,261 @@
+"""Calibration scoreboard — "dediğimiz vs olan" karnesi.
+
+Measures, on settled real-odds matches, how honest the DISPLAYED numbers are:
+
+  1. ANCHORED PIPELINE (what V35 shows): per market (MS 1/X/2, OU2.5, BTTS)
+     reliability buckets — mean stated probability vs actual frequency,
+     plus ECE / Brier per market.
+  2. SCORE CARD (V36): modal-score hit vs stated modal probability, top-5
+     coverage, HT modal hit.
+  3. STORED RUNS: prediction_runs settled per engine_version (the
+     `.sim-finished` buckets — the user's manual finished-match tests — are
+     reported separately and never mixed into the live karne).
+
+It recomputes the anchored numbers with the SAME modules the engine ships
+(models/market_anchor.py + models/score_matrix.py), so the scoreboard always
+grades current pipeline math, not a copy of it.
+
+DB: uses DATABASE_URL (data/db.py). Reads are gentle: a server-side cursor
+over an indexed, date-bounded join — never aggregate-scans the giant odds
+tables (prod runs on a Raspberry Pi).
+
+Usage:
+    python scripts/calibration_scoreboard.py [--days 365] [--buckets 10]
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+import time
+from collections import defaultdict
+from typing import Any, Dict, List, Optional, Tuple
+
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+
+import psycopg2  # noqa: E402
+from psycopg2.extras import RealDictCursor  # noqa: E402
+
+from data.db import get_clean_dsn  # noqa: E402
+from models.market_anchor import apply_corrections  # noqa: E402
+from models.score_matrix import build_calibrated_score_package  # noqa: E402
+
+REAL_ODDS_MIN_OVERROUND = 0.05  # the user's hard rule: no real odds -> excluded
+
+
+def _fetch_settled_matches(days: int) -> List[Dict[str, Any]]:
+    """Finished, real-odds matches with stored de-vigged implied probs."""
+    since_ms = int((time.time() - days * 86400) * 1000)
+    sql = """
+        SELECT f.implied_home, f.implied_draw, f.implied_away,
+               f.implied_over25, f.implied_btts_yes, f.odds_overround,
+               m.score_home, m.score_away, m.ht_score_home, m.ht_score_away
+        FROM football_ai_features f
+        JOIN matches m ON m.id = f.match_id
+        WHERE m.sport = 'football'
+          AND m.winner IN ('home', 'away', 'draw')
+          AND m.score_home IS NOT NULL
+          AND f.odds_overround > %s
+          AND m.mst_utc >= %s
+    """
+    rows: List[Dict[str, Any]] = []
+    with psycopg2.connect(get_clean_dsn()) as conn:
+        with conn.cursor() as cur:
+            cur.execute("SET statement_timeout = '120s'")
+        # server-side (named) cursor: streams gently instead of one big fetch
+        with conn.cursor("scoreboard_stream", cursor_factory=RealDictCursor) as cur:
+            cur.itersize = 5000
+            cur.execute(sql, (REAL_ODDS_MIN_OVERROUND, since_ms))
+            for r in cur:
+                rows.append(dict(r))
+    return rows
+
+
+def _anchored_probs(row: Dict[str, Any]) -> Optional[Tuple[float, float, float]]:
+    """The MS vector the V35 pipeline would display (devig is already done in
+    the stored features; apply the active home-favourite correction)."""
+    try:
+        p1 = float(row["implied_home"]); px = float(row["implied_draw"]); p2 = float(row["implied_away"])
+    except (TypeError, ValueError):
+        return None
+    if not (0.0 < p1 < 1.0 and 0.0 < px < 1.0 and 0.0 < p2 < 1.0):
+        return None
+    if abs(p1 + px + p2 - 1.0) > 0.02:  # not a clean de-vigged vector
+        return None
+    return apply_corrections(p1, px, p2)
+
+
+class Reliability:
+    """Accumulates (stated probability, outcome) pairs into buckets."""
+
+    def __init__(self, n_buckets: int) -> None:
+        self.n_buckets = n_buckets
+        self.n = defaultdict(int)
+        self.sum_p = defaultdict(float)
+        self.sum_y = defaultdict(int)
+
+    def add(self, p: float, hit: bool) -> None:
+        b = min(self.n_buckets - 1, int(p * self.n_buckets))
+        self.n[b] += 1
+        self.sum_p[b] += p
+        self.sum_y[b] += 1 if hit else 0
+
+    def report(self, title: str) -> Tuple[float, float]:
+        total = sum(self.n.values())
+        if not total:
+            print(f"\n== {title}: no data ==")
+            return 0.0, 0.0
+        ece = 0.0
+        brier_num = 0.0
+        print(f"\n== {title}  (n={total}) ==")
+        print(f"{'band':>10} {'n':>8} {'said%':>8} {'actual%':>8} {'gap_pt':>7}")
+        for b in sorted(self.n):
+            n = self.n[b]
+            said = self.sum_p[b] / n
+            act = self.sum_y[b] / n
+            ece += n * abs(said - act)
+            print(f"{b / self.n_buckets:>5.2f}-{(b + 1) / self.n_buckets:<4.2f} "
+                  f"{n:>8} {100 * said:>8.1f} {100 * act:>8.1f} {100 * (act - said):>7.1f}")
+        ece /= total
+        # Brier from bucket stats is approximate; recompute exactly elsewhere
+        # if needed. ECE is the headline honesty metric here.
+        print(f"{'ECE':>10}: {100 * ece:.2f}%")
+        return ece, brier_num
+
+
+def grade_pipeline(rows: List[Dict[str, Any]], n_buckets: int) -> None:
+    ms1 = Reliability(n_buckets); msx = Reliability(n_buckets); ms2 = Reliability(n_buckets)
+    ou = Reliability(n_buckets); btts = Reliability(n_buckets)
+    top1 = top5 = ht1 = 0
+    stated_modal = 0.0
+    n_score = 0
+
+    for r in rows:
+        anch = _anchored_probs(r)
+        sh, sa = int(r["score_home"]), int(r["score_away"])
+        winner = "home" if sh > sa else "away" if sa > sh else "draw"
+        if anch is not None:
+            p1, px, p2 = anch
+            ms1.add(p1, winner == "home")
+            msx.add(px, winner == "draw")
+            ms2.add(p2, winner == "away")
+        # exactly-0.5 values are DEFAULT FILL for matches without a real OU/BTTS
+        # market (measured: 15,993 of 78k OU rows) — never grade or use them.
+        try:
+            po = float(r["implied_over25"])
+            if po == 0.5 or not (0.05 < po < 0.95):
+                po = None
+            else:
+                ou.add(po, sh + sa >= 3)
+        except (TypeError, ValueError):
+            po = None
+        try:
+            pb = float(r["implied_btts_yes"])
+            if pb != 0.5 and 0.05 < pb < 0.95:
+                btts.add(pb, sh > 0 and sa > 0)
+        except (TypeError, ValueError):
+            pass
+
+        # V36 score card (sampled fully — pure math, no I/O)
+        if anch is not None and po is not None and 0.05 < po < 0.95:
+            pkg = build_calibrated_score_package(*anch, po)
+            actual = f"{min(sh, 10)}-{min(sa, 10)}"
+            n_score += 1
+            stated_modal += float(pkg["scenario_top5"][0]["prob"])
+            if pkg["ft"] == actual:
+                top1 += 1
+            if actual in [d["score"] for d in pkg["scenario_top5"]]:
+                top5 += 1
+            hh, ha = r.get("ht_score_home"), r.get("ht_score_away")
+            if hh is not None and ha is not None and pkg["ht"] == f"{min(int(hh),10)}-{min(int(ha),10)}":
+                ht1 += 1
+
+    ms1.report("MS ev (1) — anchored pipeline")
+    msx.report("MS beraberlik (X) — anchored pipeline")
+    ms2.report("MS deplasman (2) — anchored pipeline")
+    ou.report("Ust/Alt 2.5 (over) — devig")
+    btts.report("KG Var — devig")
+
+    if n_score:
+        print(f"\n== V36 skor karti (n={n_score}) ==")
+        print(f"  modal skor isabeti : {100 * top1 / n_score:.1f}%  (soylenen: {100 * stated_modal / n_score:.1f}%)")
+        print(f"  top-5 kapsama      : {100 * top5 / n_score:.1f}%")
+        print(f"  IY modal isabeti   : {100 * ht1 / n_score:.1f}%")
+
+
+def grade_stored_runs() -> None:
+    """Settle prediction_runs main_pick stated probabilities per engine_version.
+    `.sim-finished` buckets (manual finished-match tests) report separately."""
+    sql = """
+        SELECT pr.engine_version,
+               pr.payload_summary->'main_pick'->>'market' AS market,
+               pr.payload_summary->'main_pick'->>'pick' AS pick,
+               COALESCE((pr.payload_summary->'main_pick'->>'calibrated_probability')::float,
+                        (pr.payload_summary->'main_pick'->>'probability')::float) AS p,
+               m.score_home AS sh, m.score_away AS sa, m.winner AS w
+        FROM prediction_runs pr
+        JOIN matches m ON m.id = pr.match_id
+        WHERE m.score_home IS NOT NULL
+          AND jsonb_typeof(pr.payload_summary->'main_pick') = 'object'
+    """
+    with psycopg2.connect(get_clean_dsn()) as conn:
+        with conn.cursor() as cur:
+            cur.execute("SET statement_timeout = '60s'")
+        with conn.cursor(cursor_factory=RealDictCursor) as cur:
+            cur.execute(sql)
+            rows = cur.fetchall()
+
+    def settle(market: str, pick: str, sh: int, sa: int, w: str) -> Optional[bool]:
+        total = sh + sa
+        pick_u = (pick or "").upper()
+        over = "UST" in pick_u.replace("Ü", "U") or "OVER" in pick_u
+        if market == "MS":
+            return {"1": w == "home", "X": w == "draw", "2": w == "away"}.get(pick)
+        if market in ("OU15", "OU25", "OU35"):
+            line = {"OU15": 1.5, "OU25": 2.5, "OU35": 3.5}[market]
+            return total > line if over else total < line
+        if market == "BTTS":
+            yes = "VAR" in pick_u or "YES" in pick_u
+            return (sh > 0 and sa > 0) if yes else not (sh > 0 and sa > 0)
+        return None
+
+    stats: Dict[str, List[Tuple[float, bool]]] = defaultdict(list)
+    for r in rows:
+        if r["p"] is None:
+            continue
+        hit = settle(str(r["market"]), str(r["pick"]), int(r["sh"]), int(r["sa"]), str(r["w"]))
+        if hit is None:
+            continue
+        stats[str(r["engine_version"])].append((float(r["p"]), bool(hit)))
+
+    print("\n== prediction_runs karnesi (main_pick, soylenen vs olan) ==")
+    print(f"{'engine_version':<34} {'n':>5} {'said%':>8} {'actual%':>8}")
+    for ver in sorted(stats):
+        pairs = stats[ver]
+        n = len(pairs)
+        said = sum(p for p, _ in pairs) / n
+        act = sum(1 for _, h in pairs if h) / n
+        tag = "  <- test kovasi" if ver.endswith(".sim-finished") else ""
+        print(f"{ver:<34} {n:>5} {100 * said:>8.1f} {100 * act:>8.1f}{tag}")
+    if not stats:
+        print("  (settle edilebilir kayit yok)")
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--days", type=int, default=365, help="lookback window (days)")
+    ap.add_argument("--buckets", type=int, default=10)
+    args = ap.parse_args()
+
+    t0 = time.time()
+    rows = _fetch_settled_matches(args.days)
+    print(f"settled real-odds matches loaded: {len(rows)} (last {args.days} days, "
+          f"{time.time() - t0:.1f}s)")
+    if rows:
+        grade_pipeline(rows, args.buckets)
+    grade_stored_runs()
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/capture_closing_odds.py

@@ -0,0 +1,136 @@
+"""
+Capture Closing Odds — snapshot #2 of the minimal 2-snapshot CLV system.
+=======================================================================
+WHY: CLV (closing line value) is the only reliable proof of betting edge.
+This codebase never captured it: odds are stored as a single static snapshot
+and `odds_history` is empty. But the live sync (DataFetcherTask CRON 1) DOES
+refresh `live_matches.odds` every 15 min before kickoff, and prediction_runs
+already store the bet-time odds blob (odds_snapshot.odds, source=live_match).
+
+This script supplies the missing half: just before kickoff it copies the
+*current* live odds blob onto the match's latest prediction_run as
+`odds_snapshot.closing_odds`. Later, CLV per bet = bet-time pick odds vs
+closing pick odds (computed in live_scoreboard.py once enough data exists).
+
+Run it every ~15 min (e.g. alongside the existing sync, or its own cron):
+  python scripts/capture_closing_odds.py            # default 25-min window
+  python scripts/capture_closing_odds.py --window-min 20 --dry-run
+
+Structure-agnostic: stores the whole live odds blob; no pick parsing here.
+Idempotent: skips runs that already have closing_odds. Only ADDS a JSON key,
+never deletes. Safe to run repeatedly.
+
+⚠️ Needs one supervised test run against a live DB with upcoming matches
+   before scheduling (DB was down at authoring time).
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+import time
+from datetime import datetime, timezone
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try:
+        sys.stdout.reconfigure(encoding="utf-8")
+    except Exception:
+        pass
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+AI_ENGINE_DIR = os.path.dirname(SCRIPT_DIR)
+sys.path.insert(0, AI_ENGINE_DIR)
+
+from data.db import get_clean_dsn  # noqa: E402
+import psycopg2  # noqa: E402
+from psycopg2.extras import RealDictCursor  # noqa: E402
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--window-min", type=int, default=25,
+                    help="Capture matches kicking off within the next N minutes (default 25)")
+    ap.add_argument("--grace-min", type=int, default=10,
+                    help="Also include matches that kicked off up to N min ago (default 10)")
+    ap.add_argument("--dry-run", action="store_true",
+                    help="Report what would be captured without writing")
+    args = ap.parse_args()
+
+    now_ms = int(time.time() * 1000)
+    lo_ms = now_ms - args.grace_min * 60 * 1000
+    hi_ms = now_ms + args.window_min * 60 * 1000
+
+    captured = skipped = no_run = 0
+    with psycopg2.connect(get_clean_dsn()) as conn:
+        with conn.cursor(cursor_factory=RealDictCursor) as cur:
+            # Upcoming/just-started live matches that still hold pre-kickoff odds.
+            cur.execute(
+                """
+                SELECT id, mst_utc, odds
+                FROM live_matches
+                WHERE odds IS NOT NULL
+                  AND mst_utc BETWEEN %s AND %s
+                ORDER BY mst_utc ASC
+                """,
+                (lo_ms, hi_ms),
+            )
+            matches = cur.fetchall()
+            print(f"[capture_closing_odds] window={args.window_min}m grace={args.grace_min}m "
+                  f"upcoming_with_odds={len(matches)} dry_run={args.dry_run}")
+
+            for m in matches:
+                mid = m["id"]
+                cur.execute(
+                    """
+                    SELECT id, odds_snapshot
+                    FROM prediction_runs
+                    WHERE match_id = %s
+                    ORDER BY generated_at DESC
+                    LIMIT 1
+                    """,
+                    (mid,),
+                )
+                run = cur.fetchone()
+                if not run:
+                    no_run += 1
+                    continue
+                snap = run["odds_snapshot"] or {}
+                if isinstance(snap, str):
+                    try:
+                        snap = json.loads(snap)
+                    except Exception:
+                        snap = {}
+                if snap.get("closing_odds") is not None:
+                    skipped += 1
+                    continue
+
+                patch = {
+                    "closing_odds": m["odds"],
+                    "closing_captured_at": datetime.now(timezone.utc).isoformat(),
+                    "closing_mst_utc": m["mst_utc"],
+                    "closing_source": "live_match",
+                }
+                if args.dry_run:
+                    captured += 1
+                    print(f"  would capture match={mid} run_id={run['id']} mst_utc={m['mst_utc']}")
+                    continue
+                cur.execute(
+                    """
+                    UPDATE prediction_runs
+                    SET odds_snapshot = COALESCE(odds_snapshot, '{}'::jsonb) || %s::jsonb
+                    WHERE id = %s
+                    """,
+                    (json.dumps(patch, default=str), run["id"]),
+                )
+                captured += 1
+        if not args.dry_run:
+            conn.commit()
+
+    print(f"[capture_closing_odds] captured={captured} already_had={skipped} "
+          f"no_prediction_run={no_run}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

ai-engine/scripts/clv_report.py

@@ -0,0 +1,224 @@
+"""
+CLV Report — the single most important edge metric.
+===================================================
+Closing Line Value = did we bet at better odds than the market's closing line?
+Consistently positive CLV is the only reliable proof of a real betting edge;
+negative CLV means no edge, regardless of short-term wins/losses.
+
+This codebase stores the BET-TIME odds for ~92% of runs (prediction_runs.
+odds_snapshot.source = 'live_match' with the live odds blob, and the pick's
+odds in payload main_pick.odds). For the closing line we use, in order:
+  1. odds_snapshot.closing_odds  (captured by capture_closing_odds.py, forward)
+  2. odd_selections current value (the static near-final capture — a proxy)
+
+CLV per bet = bet_odds / closing_odds - 1   (positive = beat the close = good).
+
+Read-only. SELECT only.
+Usage:
+  python scripts/clv_report.py
+  python scripts/clv_report.py --staked-only
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from collections import defaultdict
+from typing import Any, Dict, Optional, Tuple
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try:
+        sys.stdout.reconfigure(encoding="utf-8")
+    except Exception:
+        pass
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+AI_ENGINE_DIR = os.path.dirname(SCRIPT_DIR)
+sys.path.insert(0, AI_ENGINE_DIR)
+
+from data.db import get_clean_dsn  # noqa: E402
+import psycopg2  # noqa: E402
+from psycopg2.extras import RealDictCursor  # noqa: E402
+
+# market code -> (Turkish odds-category name, pick-normalizer -> selection key)
+OU_CATS = {"OU05": "0,5 Alt/Üst", "OU15": "1,5 Alt/Üst", "OU25": "2,5 Alt/Üst",
+           "OU35": "3,5 Alt/Üst", "OU45": "4,5 Alt/Üst"}
+
+
+def _f(x: Any, d: Optional[float] = None) -> Optional[float]:
+    try:
+        return float(x) if x is not None else d
+    except (TypeError, ValueError):
+        return d
+
+
+def _parse(j: Any) -> Dict[str, Any]:
+    if isinstance(j, str):
+        try:
+            return json.loads(j)
+        except Exception:
+            return {}
+    return j or {}
+
+
+def map_pick(market: str, pick: str) -> Optional[Tuple[str, str]]:
+    """Return (category_name, selection_key) for the live-odds JSON / odd_selections."""
+    m = (market or "").upper()
+    p = (pick or "").strip()
+    pl = p.casefold()
+    if m in ("MS", "ML", "1X2"):
+        return ("Maç Sonucu", p if p in ("1", "X", "2") else None) if p in ("1", "X", "2") else None
+    if m == "HT":
+        return ("1. Yarı Sonucu", p) if p in ("1", "X", "2") else None
+    if m in OU_CATS:
+        if "üst" in pl or "ust" in pl or "over" in pl:
+            return (OU_CATS[m], "Üst")
+        if "alt" in pl or "under" in pl:
+            return (OU_CATS[m], "Alt")
+        return None
+    if m == "DC":
+        key = p.upper().replace(" ", "").replace("/", "-")
+        norm = {"1X": "1-X", "X1": "1-X", "X2": "X-2", "2X": "X-2",
+                "12": "1-2", "21": "1-2", "1-X": "1-X", "X-2": "X-2", "1-2": "1-2"}.get(key)
+        return ("Çifte Şans", norm) if norm else None
+    if m == "BTTS":
+        if "var" in pl or "yes" in pl:
+            return ("Karşılıklı Gol", "Var")
+        if "yok" in pl or "no" in pl:
+            return ("Karşılıklı Gol", "Yok")
+        return None
+    if m == "OE":
+        if "tek" in pl or "odd" in pl:
+            return ("Tek/Çift", "Tek")
+        if "çift" in pl or "cift" in pl or "even" in pl:
+            return ("Tek/Çift", "Çift")
+        return None
+    return None
+
+
+def closing_from_blob(blob: Any, cat: str, sel: str) -> Optional[float]:
+    blob = _parse(blob)
+    cat_map = blob.get(cat) if isinstance(blob, dict) else None
+    if isinstance(cat_map, dict):
+        return _f(cat_map.get(sel))
+    return None
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--staked-only", action="store_true",
+                    help="Only playable/staked bets (default: all picks with a mappable market)")
+    args = ap.parse_args()
+
+    rows_out = []
+    with psycopg2.connect(get_clean_dsn()) as conn:
+        with conn.cursor(cursor_factory=RealDictCursor) as cur:
+            cur.execute("""
+                SELECT match_id, engine_version, odds_snapshot, payload_summary,
+                       eventual_outcome, unit_profit
+                FROM prediction_runs
+                WHERE odds_snapshot->>'source' = 'live_match'
+                ORDER BY generated_at ASC
+            """)
+            runs = cur.fetchall()
+
+            for r in runs:
+                snap = _parse(r["odds_snapshot"])
+                ps = _parse(r["payload_summary"])
+                mp = ps.get("main_pick") or {}
+                market = mp.get("market")
+                pick = mp.get("pick")
+                bet_odds = _f(mp.get("odds"))
+                playable = bool(mp.get("playable"))
+                if args.staked_only and not playable:
+                    continue
+                if not market or not pick or not bet_odds or bet_odds <= 1.0:
+                    continue
+                mapped = map_pick(market, pick)
+                if not mapped or not mapped[1]:
+                    continue
+                cat, sel = mapped
+
+                # closing line: prefer captured closing_odds, else static odd_selections
+                closing = closing_from_blob(snap.get("closing_odds"), cat, sel)
+                src = "captured"
+                if closing is None:
+                    cur.execute("""
+                        SELECT os.odd_value FROM odd_categories oc
+                        JOIN odd_selections os ON os.odd_category_db_id = oc.db_id
+                        WHERE oc.match_id = %s AND oc.name = %s AND os.name = %s
+                        LIMIT 1
+                    """, (r["match_id"], cat, sel))
+                    row = cur.fetchone()
+                    closing = _f(row["odd_value"]) if row else None
+                    src = "static_proxy"
+                if closing is None or closing <= 1.0:
+                    continue
+
+                clv = bet_odds / closing - 1.0
+                rows_out.append({
+                    "market": market, "playable": playable,
+                    "bet_odds": bet_odds, "closing": closing, "clv": clv,
+                    "src": src, "profit": _f(r["unit_profit"], 0.0) or 0.0,
+                    "settled": r["eventual_outcome"] is not None
+                    and not str(r["eventual_outcome"]).startswith("NO_BET"),
+                })
+
+    if not rows_out:
+        print("No mappable runs with both bet-time and closing odds found.")
+        return 0
+
+    def agg(rs):
+        n = len(rs)
+        clvs = [x["clv"] for x in rs]
+        pos = sum(1 for c in clvs if c > 0)
+        return {
+            "n": n,
+            "mean_clv_pct": round(100.0 * sum(clvs) / n, 2),
+            "pct_positive": round(100.0 * pos / n, 1),
+            "captured": sum(1 for x in rs if x["src"] == "captured"),
+        }
+
+    print("=" * 70)
+    print("CLV REPORT  —  did we beat the closing line? (the edge compass)")
+    print("=" * 70)
+    o = agg(rows_out)
+    print(f"runs analyzed: {o['n']}   (closing source: {o['captured']} captured, "
+          f"{o['n'] - o['captured']} static-proxy)")
+    print(f"\nOVERALL mean CLV: {o['mean_clv_pct']}%   "
+          f"bets beating close: {o['pct_positive']}%")
+    print("  (positive mean CLV = real edge; ~0 or negative = no edge)\n")
+
+    staked = [x for x in rows_out if x["playable"]]
+    if staked:
+        s = agg(staked)
+        print(f"STAKED only: n={s['n']}  mean CLV={s['mean_clv_pct']}%  "
+              f"beating close={s['pct_positive']}%\n")
+
+    print("BY MARKET")
+    by_m = defaultdict(list)
+    for x in rows_out:
+        by_m[x["market"]].append(x)
+    for m, rs in sorted(by_m.items(), key=lambda kv: -len(kv[1])):
+        a = agg(rs)
+        print(f"  {m:<8} n={a['n']:>4}  mean CLV={a['mean_clv_pct']:>7}%  "
+              f"beating close={a['pct_positive']:>5}%")
+
+    # CLV vs outcome sanity: do positive-CLV bets actually win more / lose less?
+    print("\nCLV vs realized P/L (settled staked)")
+    ss = [x for x in rows_out if x["playable"] and x["settled"]]
+    if ss:
+        posc = [x for x in ss if x["clv"] > 0]
+        negc = [x for x in ss if x["clv"] <= 0]
+        for label, grp in (("CLV>0", posc), ("CLV<=0", negc)):
+            if grp:
+                pr = sum(x["profit"] for x in grp)
+                print(f"  {label:<7} n={len(grp):>3}  profit={pr:>7.2f}u  "
+                      f"ROI(flat1u)={round(100*pr/len(grp),1)}%")
+    print("=" * 70)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

ai-engine/scripts/compare_backtests.py

@@ -0,0 +1,134 @@
+"""
+Compare two diagnostic_backtest CSV outputs side-by-side.
+Used to validate that a filter change actually improved ROI vs the
+baseline run — and to detect overfitting (in-sample success but
+out-of-sample collapse).
+
+Usage:
+  python scripts/compare_backtests.py <baseline.csv> <validation.csv>
+  python scripts/compare_backtests.py  (auto-picks 2 most recent CSVs)
+"""
+
+import sys, os, glob
+import pandas as pd
+from typing import Dict
+
+REPORTS_DIR = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "reports")
+
+
+def load(path: str) -> pd.DataFrame:
+    df = pd.read_csv(path)
+    df["won_bool"] = df["won"].map(
+        {True: True, False: False, "True": True, "False": False, 1: True, 0: False}
+    )
+    return df
+
+
+def stats(df: pd.DataFrame, mask=None) -> Dict:
+    if mask is not None:
+        df = df[mask]
+    playable = df[(df["playable"] == True) & (df["won_bool"].notna())]
+    if len(playable) == 0:
+        return {"n_total": len(df), "n_playable": 0, "hit": 0, "profit": 0,
+                "staked": 0, "roi": 0}
+    wins = playable["won_bool"].sum()
+    profit = playable["unit_profit"].sum()
+    staked = playable["stake_units"].sum()
+    return {
+        "n_total": int(len(df)),
+        "n_playable": int(len(playable)),
+        "wins": int(wins),
+        "losses": int(len(playable) - wins),
+        "hit": round(100.0 * wins / len(playable), 2),
+        "profit": round(profit, 2),
+        "staked": round(staked, 2),
+        "roi": round(100.0 * profit / staked, 2) if staked else 0,
+    }
+
+
+def line(label: str, a: Dict, b: Dict, suffix: str = ""):
+    fields = ["n_total", "n_playable", "hit", "profit", "staked", "roi"]
+    parts = [f"{label:<28}"]
+    for f in fields:
+        va = a.get(f, "-")
+        vb = b.get(f, "-")
+        parts.append(f"{f}: {str(va):>8} → {str(vb):>8}")
+    print("  " + " | ".join(parts) + suffix)
+
+
+def main():
+    if len(sys.argv) == 3:
+        a_path, b_path = sys.argv[1], sys.argv[2]
+    else:
+        files = sorted(glob.glob(os.path.join(REPORTS_DIR, "diagnostic_backtest_*.csv")),
+                       key=os.path.getmtime, reverse=True)
+        if len(files) < 2:
+            print("Need at least 2 backtest CSVs in reports/. Pass paths manually.")
+            return
+        b_path, a_path = files[0], files[1]  # newest first as "validation"
+
+    print(f"Baseline    A: {os.path.basename(a_path)}")
+    print(f"Validation  B: {os.path.basename(b_path)}")
+
+    a = load(a_path)
+    b = load(b_path)
+
+    print(f"\n{'=' * 100}")
+    print(f"  OVERALL")
+    print(f"{'=' * 100}")
+    line("ALL", stats(a), stats(b))
+
+    print(f"\n{'─' * 100}")
+    print(f"  PER MARKET")
+    print(f"{'─' * 100}")
+    markets = sorted(set(a["market"].dropna().unique()) | set(b["market"].dropna().unique()))
+    for m in markets:
+        line(f"market={m}",
+             stats(a, a["market"] == m),
+             stats(b, b["market"] == m))
+
+    # New veto family check — did MUTED_MARKETS actually mute?
+    print(f"\n{'─' * 100}")
+    print(f"  NEW VETO IMPACT (look for new veto names in betting_brain.vetoes)")
+    print(f"{'─' * 100}")
+    new_vetoes = ["market_muted_by_backtest", "negative_ev_edge", "ev_edge_too_high_trap",
+                  "outside_envelope_edge_low", "outside_envelope_edge_high",
+                  "outside_envelope_odds_low", "outside_envelope_v27_must_agree"]
+    for veto in new_vetoes:
+        a_hits = a["bb_vetoes"].fillna("").str.contains(veto).sum()
+        b_hits = b["bb_vetoes"].fillna("").str.contains(veto).sum()
+        print(f"  {veto:<45}  A={a_hits:>4}  B={b_hits:>4}")
+
+    # Top issue tags
+    print(f"\n{'─' * 100}")
+    print(f"  BTTS MUTE CHECK — should be ~0 playable in validation")
+    print(f"{'─' * 100}")
+    a_btts_play = ((a["market"] == "BTTS") & (a["playable"] == True)).sum()
+    b_btts_play = ((b["market"] == "BTTS") & (b["playable"] == True)).sum()
+    print(f"  BTTS playable bets:  A={a_btts_play}  →  B={b_btts_play}  "
+          f"(should be 0 in B if MUTE works)")
+
+    # Verdict
+    print(f"\n{'=' * 100}")
+    a_s = stats(a)
+    b_s = stats(b)
+    roi_delta = b_s["roi"] - a_s["roi"]
+    if b_s["n_playable"] < 20:
+        verdict = "TOO FEW BETS — sample insufficient"
+    elif roi_delta > 5 and b_s["roi"] > 0:
+        verdict = "✅ FILTERS WORK — ROI improved AND positive"
+    elif roi_delta > 5:
+        verdict = "🟡 PARTIAL — ROI improved but still negative"
+    elif roi_delta > 0:
+        verdict = "🟡 SLIGHT IMPROVEMENT"
+    elif roi_delta < -5:
+        verdict = "❌ OVERFITTING — validation ROI collapsed"
+    else:
+        verdict = "❌ NO MATERIAL CHANGE"
+    print(f"  VERDICT: {verdict}")
+    print(f"  ROI: {a_s['roi']}% → {b_s['roi']}% (Δ {roi_delta:+.2f}pp)")
+    print(f"{'=' * 100}")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/diagnostic_backtest.py

@@ -0,0 +1,738 @@
+"""
+Diagnostic Backtest
+===================
+Run the full V28 orchestrator (in-process — no HTTP) on a window of completed
+matches, capture the recommendation + key signal features + the actual outcome,
+and produce a *diagnostic* report: not just "what was the hit rate" but
+"which feature clusters drive the losing bets".
+
+Outputs:
+  - reports/diagnostic_backtest_YYYYMMDD.csv      (per-bet detail)
+  - reports/diagnostic_backtest_YYYYMMDD.json     (aggregate metrics)
+  - reports/diagnostic_backtest_YYYYMMDD.txt      (human-readable summary)
+
+Usage:
+  python scripts/diagnostic_backtest.py --days 14 --max-matches 2000
+  python scripts/diagnostic_backtest.py --start 2026-05-10 --end 2026-05-24
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+import time
+import traceback
+from collections import defaultdict, Counter
+from datetime import datetime, timedelta, timezone
+from typing import Any, Dict, List, Optional, Tuple
+
+import psycopg2
+from psycopg2.extras import RealDictCursor
+
+# Path bootstrap so we can import the ai-engine package from anywhere
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+AI_ENGINE_DIR = os.path.dirname(SCRIPT_DIR)
+sys.path.insert(0, AI_ENGINE_DIR)
+
+from data.db import get_clean_dsn
+from services.single_match_orchestrator import get_single_match_orchestrator
+
+REPORTS_DIR = os.path.join(AI_ENGINE_DIR, "reports")
+os.makedirs(REPORTS_DIR, exist_ok=True)
+
+# Days with confirmed feeder gaps — exclude from sample
+EXCLUDED_DATES = {"2026-05-03", "2026-04-29"}
+
+
+# ── Outcome resolution ────────────────────────────────────────────────
+def _norm_pick(pick: Optional[str]) -> str:
+    return str(pick or "").strip().casefold()
+
+
+def resolve_outcome(market: str, pick: str, sh: int, sa: int,
+                    htsh: Optional[int], htsa: Optional[int]) -> Optional[bool]:
+    """Mirror of prediction-settlement.market-resolver.ts (TS side).
+    Returns True/False on settle, None if cannot resolve."""
+    m = (market or "").upper().replace(" ", "").replace("-", "_")
+    p = _norm_pick(pick)
+
+    if m in ("MS", "ML", "1X2"):
+        outcome = "1" if sh > sa else "2" if sa > sh else "x"
+        return p in {outcome, outcome.upper(), outcome.lower(), "0" if outcome == "x" else outcome}
+
+    if m in ("HT", "IY"):
+        if htsh is None or htsa is None:
+            return None
+        outcome = "1" if htsh > htsa else "2" if htsa > htsh else "x"
+        return p in {outcome, "0" if outcome == "x" else outcome}
+
+    if m in ("OU05", "OU15", "OU25", "OU35", "OU45", "TOTAL"):
+        line = {"OU05": 0.5, "OU15": 1.5, "OU25": 2.5, "OU35": 3.5,
+                "OU45": 4.5, "TOTAL": 2.5}[m]
+        total = sh + sa
+        if total == line:
+            return None
+        is_over = total > line
+        if "over" in p or "üst" in p or "ust" in p:
+            return is_over
+        if "alt" in p or "under" in p:
+            return not is_over
+        return None
+
+    if m in ("OU05_HT", "OU15_HT", "OU25_HT", "HT_OU05", "HT_OU15", "HT_OU25"):
+        if htsh is None or htsa is None:
+            return None
+        line = {"OU05_HT": 0.5, "OU15_HT": 1.5, "OU25_HT": 2.5,
+                "HT_OU05": 0.5, "HT_OU15": 1.5, "HT_OU25": 2.5}[m]
+        total = htsh + htsa
+        if total == line:
+            return None
+        is_over = total > line
+        if "over" in p or "üst" in p or "ust" in p:
+            return is_over
+        if "alt" in p or "under" in p:
+            return not is_over
+        return None
+
+    if m in ("BTTS", "KG"):
+        both = sh > 0 and sa > 0
+        if "yes" in p or "var" in p:
+            return both
+        if "no" in p or "yok" in p:
+            return not both
+        return None
+
+    if m in ("HTFT", "IYMS"):
+        if htsh is None or htsa is None or "/" not in p:
+            return None
+        ht_p, ft_p = p.split("/", 1)
+        ht_actual = "1" if htsh > htsa else "2" if htsa > htsh else "x"
+        ft_actual = "1" if sh > sa else "2" if sa > sh else "x"
+        return ht_p.strip() == ht_actual and ft_p.strip() == ft_actual
+
+    if m in ("DC", "CIFTE_SANS"):
+        ft = "1" if sh > sa else "2" if sa > sh else "X"
+        raw = p.upper().replace("-", "").replace("/", "")
+        if raw in ("1X", "X1"):
+            pair = ["1", "X"]
+        elif raw in ("X2", "2X"):
+            pair = ["X", "2"]
+        elif raw in ("12", "21"):
+            pair = ["1", "2"]
+        else:
+            return None
+        return ft in pair
+
+    if m in ("OE", "TEKCIFT"):
+        is_odd = (sh + sa) % 2 == 1
+        if "tek" in p or "odd" in p:
+            return is_odd
+        if "cift" in p or "çift" in p or "even" in p:
+            return not is_odd
+        return None
+
+    return None
+
+
+def compute_unit_profit(won: Optional[bool], stake: float, odds: Optional[float]) -> float:
+    if won is None:
+        return 0.0
+    if not won:
+        return -abs(stake) if stake else -1.0
+    if not odds or odds <= 1.0:
+        return 0.0
+    return round(stake * (odds - 1.0), 4)
+
+
+# ── Data fetch ────────────────────────────────────────────────────────
+def fetch_match_window(args) -> List[Dict]:
+    dsn = get_clean_dsn()
+    if "?schema=" in dsn:
+        dsn = dsn.split("?schema=")[0]
+
+    if args.start and args.end:
+        start = datetime.strptime(args.start, "%Y-%m-%d")
+        end = datetime.strptime(args.end, "%Y-%m-%d") + timedelta(days=1)
+    else:
+        end = datetime.now(timezone.utc).replace(tzinfo=None)
+        start = end - timedelta(days=args.days)
+
+    start_ms = int(start.timestamp() * 1000)
+    end_ms = int(end.timestamp() * 1000)
+
+    excluded = sorted(EXCLUDED_DATES)
+    excluded_clause = ""
+    if excluded:
+        ex_csv = ",".join(f"'{d}'" for d in excluded)
+        excluded_clause = (
+            f" AND to_timestamp(mst_utc/1000)::date "
+            f"NOT IN ({ex_csv})"
+        )
+
+    with psycopg2.connect(dsn) as conn:
+        with conn.cursor(cursor_factory=RealDictCursor) as cur:
+            cur.execute(
+                f"""
+                SELECT id AS match_id,
+                       score_home, score_away,
+                       ht_score_home, ht_score_away,
+                       league_id,
+                       to_timestamp(mst_utc/1000)::date AS match_date
+                FROM matches
+                WHERE sport='football'
+                  AND status='FT'
+                  AND score_home IS NOT NULL
+                  AND score_away IS NOT NULL
+                  AND mst_utc >= %s
+                  AND mst_utc < %s
+                  {excluded_clause}
+                ORDER BY mst_utc DESC
+                LIMIT %s
+                """,
+                (start_ms, end_ms, args.max_matches),
+            )
+            return cur.fetchall()
+
+
+# ── Per-bet capture ───────────────────────────────────────────────────
+def capture_bet_row(match: Dict, package: Dict) -> Dict[str, Any]:
+    """Distill orchestrator response + ground truth into one analytic row."""
+    main = package.get("main_pick") or {}
+    bb = main.get("betting_brain") or {}
+    advice = package.get("bet_advice") or {}
+    v27 = package.get("v27_engine") or {}
+    triple = (v27.get("triple_value") or {})
+    risk = package.get("risk") or {}
+    quality = package.get("data_quality") or {}
+    htft_payload = ((package.get("market_board") or {}).get("HTFT") or {})
+    htft_probs = htft_payload.get("probs") or {}
+
+    sh, sa = match["score_home"], match["score_away"]
+    htsh, htsa = match["ht_score_home"], match["ht_score_away"]
+
+    market = main.get("market")
+    pick = main.get("pick")
+    odds_val = _f(main.get("odds"))
+    stake = _f(main.get("stake_units"), 1.0)
+    playable = bool(main.get("playable")) and bool(advice.get("playable"))
+
+    won = resolve_outcome(market, pick, sh, sa, htsh, htsa) if market and pick else None
+    profit = compute_unit_profit(won, stake, odds_val) if playable else 0.0
+
+    # Reversal context (only meaningful for MS picks)
+    rev_prob = None
+    if market == "MS" and pick in ("1", "2"):
+        if pick == "1":
+            rev_prob = _f(htft_probs.get("1/2"), 0.0) + _f(htft_probs.get("1/X"), 0.0)
+        else:
+            rev_prob = _f(htft_probs.get("2/1"), 0.0) + _f(htft_probs.get("2/X"), 0.0)
+
+    return {
+        "match_id": match["match_id"],
+        "match_date": str(match["match_date"]),
+        "league_id": match.get("league_id"),
+        "score_home": sh,
+        "score_away": sa,
+        "ht_score_home": htsh,
+        "ht_score_away": htsa,
+        "market": market,
+        "pick": pick,
+        "odds": odds_val,
+        "stake_units": stake,
+        "playable": playable,
+        "won": won,
+        "unit_profit": profit,
+        "raw_confidence": _f(main.get("raw_confidence")),
+        "calibrated_confidence": _f(main.get("calibrated_confidence")),
+        "play_score": _f(main.get("play_score")),
+        "ev_edge": _f(main.get("ev_edge")),
+        "bet_grade": main.get("bet_grade"),
+        "is_value_sniper": bool(main.get("is_value_sniper")),
+        "bb_score": _f(bb.get("score")),
+        "bb_action": bb.get("action"),
+        "bb_vetoes": ";".join(bb.get("vetoes") or []),
+        "bb_issues": ";".join(bb.get("issues") or []),
+        "bb_positives": ";".join(bb.get("positives") or []),
+        "bb_model_prob": _f(bb.get("model_prob")),
+        "bb_implied_prob": _f(bb.get("implied_prob")),
+        "bb_model_market_gap": _f(bb.get("model_market_gap")),
+        "bb_divergence": _f(bb.get("divergence")),
+        "bb_trap_market": bool(bb.get("trap_market_flag")),
+        "v27_consensus": v27.get("consensus"),
+        "data_quality_score": _f(quality.get("score")),
+        "data_quality_flags": ";".join(quality.get("flags") or []),
+        "risk_level": (risk.get("level") if isinstance(risk, dict) else None),
+        "odds_reliability": _f(main.get("odds_reliability")),
+        "htft_reversal_prob": rev_prob,
+        "htft_top_pick": _argmax(htft_probs),
+        "league_name": (package.get("match_info") or {}).get("league_name"),
+        "is_cup": _is_cup((package.get("match_info") or {}).get("league_name") or ""),
+        "model_version": package.get("model_version"),
+        "decision_reason": main.get("pick_reason") or advice.get("reason"),
+    }
+
+
+def _f(x: Any, default: Optional[float] = None) -> Optional[float]:
+    try:
+        return float(x) if x is not None else default
+    except (TypeError, ValueError):
+        return default
+
+
+def _argmax(d: Dict[str, Any]) -> Optional[str]:
+    best, val = None, -1.0
+    for k, v in d.items():
+        fv = _f(v, 0.0) or 0.0
+        if fv > val:
+            best, val = k, fv
+    return best
+
+
+_CUP_KEYWORDS = ("kupa", "cup", "coupe", "copa", "coppa", "pokal", "trophy",
+                 "shield", "ziraat", "süper kupa", "super cup", "beker", "taça", "taca")
+
+
+def _is_cup(name: str) -> bool:
+    n = (name or "").lower()
+    return any(kw in n for kw in _CUP_KEYWORDS)
+
+
+# ── Aggregation helpers ────────────────────────────────────────────────
+def _bucket(value: Optional[float], edges: List[float]) -> Optional[str]:
+    if value is None:
+        return None
+    for i, edge in enumerate(edges):
+        if value < edge:
+            if i == 0:
+                return f"<{edge}"
+            return f"{edges[i-1]}-{edge}"
+    return f">={edges[-1]}"
+
+
+def _summary_stats(rows: List[Dict]) -> Dict[str, Any]:
+    if not rows:
+        return {"n": 0}
+    settled = [r for r in rows if r["playable"] and r["won"] is not None]
+    won = sum(1 for r in settled if r["won"])
+    lost = sum(1 for r in settled if not r["won"])
+    profit = sum(float(r["unit_profit"]) for r in settled)
+    staked = sum(float(r["stake_units"]) for r in settled)
+    return {
+        "n_total": len(rows),
+        "n_playable_settled": len(settled),
+        "wins": won,
+        "losses": lost,
+        "hit_rate_pct": round(100.0 * won / len(settled), 2) if settled else None,
+        "unit_profit": round(profit, 3),
+        "staked": round(staked, 3),
+        "roi_pct": round(100.0 * profit / staked, 2) if staked else None,
+    }
+
+
+def aggregate(rows: List[Dict]) -> Dict[str, Any]:
+    out: Dict[str, Any] = {"overall": _summary_stats(rows)}
+
+    by = lambda key_fn: defaultdict(list)
+    market_buckets = by(None)
+    conf_buckets = by(None)
+    odds_buckets = by(None)
+    grade_buckets = by(None)
+    cup_buckets = by(None)
+    motivation_buckets = by(None)
+
+    for r in rows:
+        if r["playable"]:
+            market_buckets[r["market"] or "?"].append(r)
+            conf_buckets[_bucket(r["calibrated_confidence"],
+                                  [45, 50, 55, 60, 65, 70, 80])].append(r)
+            odds_buckets[_bucket(r["odds"], [1.3, 1.5, 1.8, 2.2, 3.0, 5.0])].append(r)
+            grade_buckets[r["bet_grade"] or "?"].append(r)
+            cup_buckets["cup" if r["is_cup"] else "league"].append(r)
+
+    out["by_market"] = {k: _summary_stats(v) for k, v in market_buckets.items()}
+    out["by_confidence"] = {k: _summary_stats(v) for k, v in conf_buckets.items() if k}
+    out["by_odds"] = {k: _summary_stats(v) for k, v in odds_buckets.items() if k}
+    out["by_grade"] = {k: _summary_stats(v) for k, v in grade_buckets.items()}
+    out["by_competition"] = {k: _summary_stats(v) for k, v in cup_buckets.items()}
+    return out
+
+
+def loss_diagnostics(rows: List[Dict]) -> Dict[str, Any]:
+    losses = [r for r in rows if r["playable"] and r["won"] is False]
+    if not losses:
+        return {"n_losses": 0}
+    n = len(losses)
+
+    def share(predicate) -> Tuple[int, float]:
+        c = sum(1 for r in losses if predicate(r))
+        return c, round(100.0 * c / n, 2)
+
+    diagnostics = {
+        "n_losses": n,
+        "total_loss_units": round(sum(float(r["unit_profit"]) for r in losses), 3),
+        "patterns": {
+            "high_htft_reversal_prob (>=0.20)": share(
+                lambda r: (r.get("htft_reversal_prob") or 0) >= 0.20
+            ),
+            "cup_match": share(lambda r: r["is_cup"]),
+            "low_league_reliability (<0.45)": share(
+                lambda r: (r.get("odds_reliability") or 1) < 0.45
+            ),
+            "v27_disagree": share(lambda r: r.get("v27_consensus") == "DISAGREE"),
+            "trap_market_flagged": share(lambda r: r.get("bb_trap_market")),
+            "low_calibrated_conf (<55)": share(
+                lambda r: (r.get("calibrated_confidence") or 0) < 55
+            ),
+            "high_odds_underdog (>=2.5)": share(
+                lambda r: (r.get("odds") or 0) >= 2.5
+            ),
+            "low_data_quality (<0.55)": share(
+                lambda r: (r.get("data_quality_score") or 1) < 0.55
+            ),
+            "high_risk_level": share(
+                lambda r: r.get("risk_level") in ("HIGH", "EXTREME")
+            ),
+            "inferred_features": share(
+                lambda r: "ai_features_inferred_from_history" in (r.get("data_quality_flags") or "")
+            ),
+        },
+        "by_market": Counter(r["market"] for r in losses).most_common(),
+        "by_league": Counter(r.get("league_name") for r in losses).most_common(10),
+    }
+
+    # Top issue tags from betting_brain across losses
+    issue_counter = Counter()
+    veto_counter = Counter()
+    for r in losses:
+        for tag in (r.get("bb_issues") or "").split(";"):
+            if tag:
+                issue_counter[tag] += 1
+        for tag in (r.get("bb_vetoes") or "").split(";"):
+            if tag:
+                veto_counter[tag] += 1
+    diagnostics["top_bb_issues_in_losses"] = issue_counter.most_common(15)
+    diagnostics["top_bb_vetoes_in_losses"] = veto_counter.most_common(15)
+    return diagnostics
+
+
+# ── Recommendations ────────────────────────────────────────────────────
+def make_recommendations(rows: List[Dict], agg: Dict[str, Any],
+                          diag: Dict[str, Any]) -> List[Dict[str, Any]]:
+    recs: List[Dict[str, Any]] = []
+    overall = agg.get("overall") or {}
+    if not overall.get("n_playable_settled"):
+        return recs
+
+    # Cross-reference market hit rate vs overall — flag chronic losers.
+    overall_hit = overall.get("hit_rate_pct") or 0.0
+    for market, stats in (agg.get("by_market") or {}).items():
+        n = stats.get("n_playable_settled") or 0
+        hit = stats.get("hit_rate_pct")
+        roi = stats.get("roi_pct")
+        if n < 30:
+            continue
+        if hit is not None and roi is not None and roi < -10 and hit < overall_hit - 10:
+            recs.append({
+                "type": "drop_market",
+                "market": market,
+                "evidence": f"hit={hit}%, roi={roi}%, n={n} — chronic loser",
+                "suggested_fix": f"Add veto in betting_brain when market=={market} unless overwhelming evidence",
+                "estimated_loss_prevented_units": round(-(stats.get("unit_profit") or 0), 2),
+            })
+
+    # Confidence band tuning — flag bands where ROI < 0 despite passing eşik
+    for band, stats in (agg.get("by_confidence") or {}).items():
+        n = stats.get("n_playable_settled") or 0
+        roi = stats.get("roi_pct")
+        if n >= 40 and roi is not None and roi < -8:
+            recs.append({
+                "type": "raise_confidence_threshold",
+                "confidence_band": band,
+                "evidence": f"n={n}, roi={roi}%",
+                "suggested_fix": f"Raise MIN_BET_SCORE or market_min_conf above {band.split('-')[0]}",
+            })
+
+    # Loss diagnostic — if cup matches dominate losses, recommend cup-aware filter
+    patterns = (diag.get("patterns") or {})
+    cup_share = patterns.get("cup_match", (0, 0))[1]
+    if cup_share >= 25:
+        recs.append({
+            "type": "cup_match_filter",
+            "evidence": f"{cup_share}% of losses are cup matches",
+            "suggested_fix": "Tighten betting_brain thresholds for is_cup_match=True picks",
+        })
+
+    rev_share = patterns.get("high_htft_reversal_prob (>=0.20)", (0, 0))[1]
+    if rev_share >= 15:
+        recs.append({
+            "type": "tighten_reversal_check",
+            "evidence": f"{rev_share}% of losses had HTFT reversal prob >=0.20 (already partial fix)",
+            "suggested_fix": "Lower reversal threshold in betting_brain from 0.25 to 0.20 for veto trigger",
+        })
+
+    rel_share = patterns.get("low_league_reliability (<0.45)", (0, 0))[1]
+    if rel_share >= 20:
+        recs.append({
+            "type": "league_reliability_filter",
+            "evidence": f"{rel_share}% of losses in low-reliability leagues (<0.45)",
+            "suggested_fix": "Add hard veto when odds_reliability<0.45 for non-value-sniper picks",
+        })
+
+    return recs
+
+
+# ── CSV / report writers ───────────────────────────────────────────────
+def write_csv(rows: List[Dict], path: str):
+    if not rows:
+        return
+    import csv
+    fields = list(rows[0].keys())
+    with open(path, "w", newline="", encoding="utf-8") as f:
+        w = csv.DictWriter(f, fieldnames=fields)
+        w.writeheader()
+        for r in rows:
+            w.writerow(r)
+
+
+def write_text_summary(rows: List[Dict], agg: Dict, diag: Dict,
+                        recs: List[Dict], path: str, args):
+    lines: List[str] = []
+    push = lines.append
+    push("=" * 78)
+    push("DIAGNOSTIC BACKTEST REPORT")
+    push("=" * 78)
+    push(f"Generated:      {datetime.now().isoformat(timespec='seconds')}")
+    push(f"Sample window:  start={args.start or f'-{args.days}d'}, end={args.end or 'now'}")
+    push(f"Max matches:    {args.max_matches}")
+    push(f"Excluded days:  {sorted(EXCLUDED_DATES)}")
+    push("")
+    push("OVERALL")
+    push("-" * 78)
+    overall = agg.get("overall") or {}
+    for k in ("n_total", "n_playable_settled", "wins", "losses",
+              "hit_rate_pct", "unit_profit", "staked", "roi_pct"):
+        push(f"  {k:25}: {overall.get(k)}")
+    push("")
+    push("PER MARKET")
+    push("-" * 78)
+    push(f"  {'market':<8} {'n':>6} {'hit%':>7} {'profit':>9} {'roi%':>7}")
+    for market, s in sorted((agg.get("by_market") or {}).items(),
+                             key=lambda kv: -(kv[1].get("n_playable_settled") or 0)):
+        push(f"  {market:<8} {s.get('n_playable_settled',0):>6} "
+             f"{str(s.get('hit_rate_pct','')):>7} "
+             f"{str(s.get('unit_profit','')):>9} "
+             f"{str(s.get('roi_pct','')):>7}")
+    push("")
+    push("PER CALIBRATED CONFIDENCE BAND")
+    push("-" * 78)
+    push(f"  {'band':<10} {'n':>6} {'hit%':>7} {'roi%':>7}")
+    for band, s in sorted((agg.get("by_confidence") or {}).items()):
+        push(f"  {band:<10} {s.get('n_playable_settled',0):>6} "
+             f"{str(s.get('hit_rate_pct','')):>7} "
+             f"{str(s.get('roi_pct','')):>7}")
+    push("")
+    push("PER ODDS BAND")
+    push("-" * 78)
+    push(f"  {'band':<10} {'n':>6} {'hit%':>7} {'roi%':>7}")
+    for band, s in sorted((agg.get("by_odds") or {}).items()):
+        push(f"  {band:<10} {s.get('n_playable_settled',0):>6} "
+             f"{str(s.get('hit_rate_pct','')):>7} "
+             f"{str(s.get('roi_pct','')):>7}")
+    push("")
+    push("LEAGUE vs CUP")
+    push("-" * 78)
+    for k, s in (agg.get("by_competition") or {}).items():
+        push(f"  {k:<8} n={s.get('n_playable_settled',0):>4} "
+             f"hit={s.get('hit_rate_pct','-')}% roi={s.get('roi_pct','-')}%")
+    push("")
+    push("LOSS DIAGNOSTICS")
+    push("-" * 78)
+    push(f"  total losses:        {diag.get('n_losses')}")
+    push(f"  total lost units:    {diag.get('total_loss_units')}")
+    push(f"  By market:           {diag.get('by_market')}")
+    push("  Loss patterns (count, % of losses):")
+    for pattern, (c, pct) in (diag.get("patterns") or {}).items():
+        push(f"    {pattern:<55} {c:>4} ({pct}%)")
+    push("  Top betting_brain issues seen in losses:")
+    for issue, c in (diag.get("top_bb_issues_in_losses") or []):
+        push(f"    {issue:<55} {c}")
+    push("  Top betting_brain vetoes (in losses — i.e. veto fired but bet still went through value-sniper override):")
+    for veto, c in (diag.get("top_bb_vetoes_in_losses") or []):
+        push(f"    {veto:<55} {c}")
+    push("")
+    push("RECOMMENDATIONS")
+    push("-" * 78)
+    if not recs:
+        push("  (none surfaced — sample too small or no clear pattern)")
+    for r in recs:
+        push(f"  • [{r['type']}]")
+        for k, v in r.items():
+            if k == "type":
+                continue
+            push(f"      {k}: {v}")
+    push("")
+    push("=" * 78)
+    with open(path, "w", encoding="utf-8") as f:
+        f.write("\n".join(lines))
+
+
+# ── Main loop ─────────────────────────────────────────────────────────
+def _checkpoint_paths(args) -> Tuple[str, str]:
+    """Stable checkpoint paths derived from the run's date window so a
+    re-run with the same args picks up the same checkpoint."""
+    key = f"{args.start or 'd' + str(args.days)}_{args.end or 'now'}_{args.max_matches}"
+    key = key.replace("-", "").replace(":", "")
+    ckpt_csv = os.path.join(REPORTS_DIR, f"_checkpoint_{key}.csv")
+    ckpt_state = os.path.join(REPORTS_DIR, f"_checkpoint_{key}.state")
+    return ckpt_csv, ckpt_state
+
+
+def _load_checkpoint(args) -> Tuple[List[Dict], set]:
+    """Read partial CSV + processed-IDs set if a previous run was interrupted."""
+    ckpt_csv, _ = _checkpoint_paths(args)
+    if not os.path.exists(ckpt_csv):
+        return [], set()
+    import csv
+    rows: List[Dict] = []
+    seen: set = set()
+    try:
+        with open(ckpt_csv, "r", encoding="utf-8", newline="") as f:
+            reader = csv.DictReader(f)
+            for row in reader:
+                rows.append(row)
+                seen.add(str(row.get("match_id") or ""))
+    except Exception as e:
+        print(f"  checkpoint read failed ({e}); starting fresh")
+        return [], set()
+    return rows, seen
+
+
+def _flush_checkpoint(args, rows: List[Dict]) -> None:
+    """Atomic-ish overwrite of the partial CSV. Cheap enough at every 100 rows."""
+    if not rows:
+        return
+    ckpt_csv, _ = _checkpoint_paths(args)
+    import csv
+    tmp = ckpt_csv + ".tmp"
+    fields = list(rows[0].keys())
+    with open(tmp, "w", encoding="utf-8", newline="") as f:
+        w = csv.DictWriter(f, fieldnames=fields)
+        w.writeheader()
+        for r in rows:
+            w.writerow(r)
+    os.replace(tmp, ckpt_csv)
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--days", type=int, default=14,
+                        help="Backwards window from now (default 14)")
+    parser.add_argument("--max-matches", type=int, default=2000,
+                        help="Hard cap on matches processed (default 2000)")
+    parser.add_argument("--start", help="Start date YYYY-MM-DD (overrides --days)")
+    parser.add_argument("--end", help="End date YYYY-MM-DD")
+    parser.add_argument("--progress-interval", type=int, default=50)
+    parser.add_argument("--checkpoint-every", type=int, default=100,
+                        help="Flush partial CSV every N matches (default 100)")
+    parser.add_argument("--no-resume", action="store_true",
+                        help="Ignore any prior checkpoint and start fresh")
+    args = parser.parse_args()
+
+    print("=" * 70)
+    print("DIAGNOSTIC BACKTEST")
+    print("=" * 70)
+    print(f"Loading orchestrator...")
+    orch = get_single_match_orchestrator()
+    # Warm V25 + V27 + basketball loaders so the first match doesn't pay it
+    try:
+        orch._get_v25_predictor()
+    except Exception as e:
+        print(f"  v25 warmup: {e}")
+    try:
+        orch._get_v27_predictor()
+    except Exception as e:
+        print(f"  v27 warmup: {e}")
+
+    print(f"Fetching match window...")
+    matches = fetch_match_window(args)
+    n = len(matches)
+    print(f"  {n} matches selected")
+
+    if not matches:
+        print("No matches to process. Exiting.")
+        return
+
+    # ── Resume from prior checkpoint if available ──
+    rows: List[Dict[str, Any]] = []
+    seen_ids: set = set()
+    if not args.no_resume:
+        rows, seen_ids = _load_checkpoint(args)
+        if rows:
+            print(f"  Resuming from checkpoint: {len(rows)} matches already done")
+    errors: List[Tuple[str, str]] = []
+    t0 = time.time()
+
+    for i, m in enumerate(matches, start=1):
+        mid = str(m["match_id"])
+        if mid in seen_ids:
+            continue
+        try:
+            pkg = orch.analyze_match(mid)
+            if pkg is None:
+                continue
+            row = capture_bet_row(m, pkg)
+            rows.append(row)
+        except KeyboardInterrupt:
+            print("\nInterrupted, flushing checkpoint...")
+            _flush_checkpoint(args, rows)
+            break
+        except Exception as e:
+            errors.append((mid, str(e)))
+            if len(errors) <= 5:
+                traceback.print_exc()
+
+        # ── Periodic checkpoint flush so a crash doesn't lose everything ──
+        if i % args.checkpoint_every == 0:
+            _flush_checkpoint(args, rows)
+
+        if i % args.progress_interval == 0:
+            elapsed = time.time() - t0
+            rate = i / elapsed
+            eta = (n - i) / rate if rate else 0
+            playable_so_far = sum(1 for r in rows if r["playable"])
+            print(f"  [{i}/{n}] rate={rate:.1f}/s eta={eta/60:.1f}min "
+                  f"playable={playable_so_far} errors={len(errors)} "
+                  f"(checkpoint at every {args.checkpoint_every})")
+
+    print(f"\nProcessed {len(rows)} rows in {(time.time()-t0):.1f}s "
+          f"({len(errors)} errors)")
+
+    # Aggregate
+    print("Aggregating...")
+    agg = aggregate(rows)
+    diag = loss_diagnostics(rows)
+    recs = make_recommendations(rows, agg, diag)
+
+    stamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    csv_path = os.path.join(REPORTS_DIR, f"diagnostic_backtest_{stamp}.csv")
+    json_path = os.path.join(REPORTS_DIR, f"diagnostic_backtest_{stamp}.json")
+    txt_path = os.path.join(REPORTS_DIR, f"diagnostic_backtest_{stamp}.txt")
+
+    write_csv(rows, csv_path)
+    with open(json_path, "w", encoding="utf-8") as f:
+        json.dump({"args": vars(args), "aggregate": agg, "loss_diagnostics": diag,
+                   "recommendations": recs, "errors_sample": errors[:20]},
+                  f, indent=2, default=str)
+    write_text_summary(rows, agg, diag, recs, txt_path, args)
+
+    print(f"\nOutputs:")
+    print(f"  CSV:  {csv_path}")
+    print(f"  JSON: {json_path}")
+    print(f"  TXT:  {txt_path}")
+    print("\nOverall:", agg.get("overall"))
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/edge_search.py

@@ -0,0 +1,181 @@
+"""
+Edge Search — is there a profitable POCKET (by league) the global model misses?
+==============================================================================
+Global leak-free MS is ~-5.6% (the vig). But efficiency varies: obscure / low-
+tier leagues may be mispriced. This walks a leak-free model forward and slices
+the value-bet ROI BY LEAGUE, requiring a real sample AND multi-fold consistency
+so we don't chase one lucky window.
+
+Leak-free: drops the confirmed/suspected leakage columns (see LEAKY). Uses odds
+in features (realistic). Value bet = biggest model_prob - implied edge > margin.
+
+⚠️ Even a positive pocket here is a LEAD, not proof: the CSV odds are a static
+capture, not the verified closing line. Anything flagged must be forward-
+validated with real CLV (capture_closing_odds.py) before staking.
+
+Usage: python scripts/edge_search.py --folds 6 --min-bets 150
+"""
+from __future__ import annotations
+import argparse, os, sys, time
+import numpy as np, pandas as pd, xgboost as xgb
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.insert(0, AI_DIR)
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+
+META = {"match_id","home_team_id","away_team_id","league_id","mst_utc",
+        "score_home","score_away","ht_score_home","ht_score_away"}
+LEAKY = {"home_goals_form","away_goals_form","total_goals","ht_total_goals",
+         "squad_diff","home_squad_quality","away_squad_quality",
+         "referee_home_bias","referee_avg_goals"}
+
+
+def league_names(ids):
+    """Resilient id->name lookup."""
+    from data.db import get_clean_dsn
+    import psycopg2
+    from psycopg2.extras import RealDictCursor
+    out = {}
+    ids = [str(i) for i in ids if i is not None]
+    if not ids: return out
+    for _ in range(3):
+        try:
+            with psycopg2.connect(get_clean_dsn()) as c:
+                with c.cursor(cursor_factory=RealDictCursor) as cur:
+                    cur.execute("SELECT id, name FROM leagues WHERE id = ANY(%s)", (ids,))
+                    for r in cur.fetchall(): out[str(r["id"])] = r["name"]
+            return out
+        except Exception:
+            time.sleep(1.0)
+    return out
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--folds", type=int, default=6)
+    ap.add_argument("--estimators", type=int, default=200)
+    ap.add_argument("--margin", type=float, default=0.0)
+    ap.add_argument("--min-bets", type=int, default=150)
+    args = ap.parse_args()
+
+    print(f"Loading {CSV} ...")
+    df = pd.read_csv(CSV, low_memory=False).sort_values("mst_utc").reset_index(drop=True)
+    sh = pd.to_numeric(df["score_home"], errors="coerce")
+    sa = pd.to_numeric(df["score_away"], errors="coerce")
+    ok = sh.notna() & sa.notna()
+    df, sh, sa = df[ok].reset_index(drop=True), sh[ok.values].values, sa[ok.values].values
+    y = np.where(sh > sa, 0, np.where(sh == sa, 1, 2))
+    league = df["league_id"].astype(str).values
+    odds = df[["odds_ms_h","odds_ms_d","odds_ms_a"]].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+
+    feats = [c for c in df.columns if c not in META and not c.startswith("label_") and c not in LEAKY]
+    X = df[feats].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    rel = pd.to_numeric(df.get("league_reliability_score", pd.Series([np.nan]*len(df))),
+                        errors="coerce").fillna(-1.0).values
+    print(f"  {len(df):,} rows  features={len(feats)} (leak-free)  folds={args.folds}")
+
+    n = len(df); start = int(n * 0.5)
+    bounds = np.linspace(start, n, args.folds + 1, dtype=int)
+    params = {"objective":"multi:softprob","num_class":3,"max_depth":5,"eta":0.05,
+              "subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+
+    # reliability quartile edges from the betting universe (rel>=0)
+    rv = rel[rel >= 0]
+    qs = np.quantile(rv, [0.25, 0.5, 0.75]) if len(rv) else [0.3, 0.5, 0.7]
+    def rel_band(x):
+        if x < 0: return "rel:unknown"
+        if x < qs[0]: return f"rel:Q1(<{qs[0]:.2f})"
+        if x < qs[1]: return f"rel:Q2"
+        if x < qs[2]: return f"rel:Q3"
+        return f"rel:Q4(>={qs[2]:.2f})"
+    def odds_band(o):
+        return ("<1.5" if o<1.5 else "1.5-2" if o<2 else "2-3" if o<3 else
+                "3-5" if o<5 else "5-8" if o<8 else "8+")
+
+    recs = []  # (group_key, fold, pnl, win)
+    glob = {"n":0,"pnl":0.0,"win":0}
+    for fi in range(args.folds):
+        te0, te1 = bounds[fi], bounds[fi+1]
+        if te1-te0 < 50: continue
+        bst = xgb.train(params, xgb.DMatrix(X[:te0], label=y[:te0]), num_boost_round=args.estimators)
+        proba = bst.predict(xgb.DMatrix(X[te0:te1]))
+        yte, ote, rte = y[te0:te1], odds[te0:te1], rel[te0:te1]
+        implied = np.where(ote > 1.0, 1.0/ote, np.nan)
+        edge = np.where(np.isnan(implied), -9.0, proba - implied)
+        pick = edge.argmax(1)
+        bet = edge[np.arange(len(yte)), pick] > args.margin
+        win = (pick == yte) & bet
+        pick_odds = ote[np.arange(len(yte)), pick]
+        pnl = np.where(win, pick_odds-1.0, -1.0)
+        for i in range(len(yte)):
+            if not bet[i]: continue
+            glob["n"]+=1; glob["pnl"]+=pnl[i]; glob["win"]+=int(win[i])
+            recs.append((rel_band(rte[i]), fi, pnl[i], int(win[i])))
+            recs.append((odds_band(pick_odds[i]), fi, pnl[i], int(win[i])))
+            recs.append((rel_band(rte[i])+" x "+odds_band(pick_odds[i]), fi, pnl[i], int(win[i])))
+        print(f"  fold {fi}: tested {len(yte):,}  bets {int(bet.sum()):,}")
+
+    print("\n"+"="*78)
+    print(f"GLOBAL leak-free: bets={glob['n']:,}  hit={100*glob['win']/max(glob['n'],1):.1f}%  "
+          f"ROI(flat1u)={100*glob['pnl']/max(glob['n'],1):.2f}%")
+    print("="*78)
+
+    rdf = pd.DataFrame(recs, columns=["grp","fold","pnl","win"])
+    def report(prefix, title):
+        sub = rdf[rdf["grp"].str.startswith(prefix)]
+        if sub.empty: return
+        print(f"\n{title}")
+        print(f"  {'bucket':<26}{'bets':>6}{'hit%':>7}{'ROI%':>8}{'folds+':>8}")
+        print("  "+"-"*54)
+        g = sub.groupby("grp")
+        out=[]
+        for k,d in g:
+            nb=len(d)
+            if nb < args.min_bets: continue
+            roi=100*d["pnl"].sum()/nb; hit=100*d["win"].sum()/nb
+            fp=d.groupby("fold")["pnl"].sum(); folds_pos=int((fp>0).sum()); ft=fp.shape[0]
+            out.append((roi,k,nb,hit,folds_pos,ft))
+        for roi,k,nb,hit,fp,ft in sorted(out,reverse=True):
+            print(f"  {k:<26}{nb:>6}{hit:>7.1f}{roi:>8.1f}{str(fp)+'/'+str(ft):>8}")
+    report("rel:", "BY LEAGUE-RELIABILITY BAND  (Q1=most obscure ... Q4=most reliable)")
+    report(("<","1","2","3","5","8"), None)  # odds bands start with digit/<
+    # odds-band buckets begin with a digit or '<'
+    sub = rdf[~rdf["grp"].str.startswith("rel:")]
+    sub = sub[~sub["grp"].str.contains(" x ")]
+    if not sub.empty:
+        print("\nBY ODDS BAND")
+        print(f"  {'bucket':<26}{'bets':>6}{'hit%':>7}{'ROI%':>8}{'folds+':>8}")
+        print("  "+"-"*54)
+        out=[]
+        for k,d in sub.groupby("grp"):
+            nb=len(d)
+            if nb<args.min_bets: continue
+            roi=100*d["pnl"].sum()/nb; hit=100*d["win"].sum()/nb
+            fp=d.groupby("fold")["pnl"].sum(); out.append((roi,k,nb,hit,int((fp>0).sum()),fp.shape[0]))
+        for roi,k,nb,hit,fpv,ft in sorted(out,reverse=True):
+            print(f"  {k:<26}{nb:>6}{hit:>7.1f}{roi:>8.1f}{str(fpv)+'/'+str(ft):>8}")
+    # 2D reliability x odds
+    sub2 = rdf[rdf["grp"].str.contains(" x ")]
+    if not sub2.empty:
+        print("\nBY RELIABILITY x ODDS  (candidate pockets, n>=min-bets)")
+        print(f"  {'bucket':<26}{'bets':>6}{'hit%':>7}{'ROI%':>8}{'folds+':>8}")
+        print("  "+"-"*54)
+        out=[]
+        for k,d in sub2.groupby("grp"):
+            nb=len(d)
+            if nb<args.min_bets: continue
+            roi=100*d["pnl"].sum()/nb; hit=100*d["win"].sum()/nb
+            fp=d.groupby("fold")["pnl"].sum(); out.append((roi,k,nb,hit,int((fp>0).sum()),fp.shape[0]))
+        for roi,k,nb,hit,fpv,ft in sorted(out,reverse=True)[:15]:
+            print(f"  {k:<26}{nb:>6}{hit:>7.1f}{roi:>8.1f}{str(fpv)+'/'+str(ft):>8}")
+    print("\nREAD: a pocket is a real LEAD only if ROI>0 AND positive in MOST folds")
+    print("(folds+ near full) AND bets large. +ROI in 1-2 folds = noise / overfit.")
+    print("Then forward-validate with CLV (capture_closing_odds.py) before staking.")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/extract_upcoming_features.py

@@ -0,0 +1,198 @@
+"""
+Extract Upcoming Features — leak-free feature rows for UPCOMING (NS) matches,
+produced by the EXACT same pipeline that built training_data_v27.csv.
+=============================================================================
+Why this exists: the picker (generate_daily_picks.py) needs the 133 leak-free
+features for tomorrow's matches, computed IDENTICALLY to training (any drift =
+train/serve skew = broken model). So we reuse V27Loader + V27Extractor verbatim:
+
+  1. load_all() builds ELO / team history / league / squad caches from FT
+     matches ONLY (untouched — guarantees identical feature computation).
+  2. We then APPEND upcoming NS matches as targets and inject their odds from
+     live_matches.odds (all markets, same mapping as the trainer's _load_odds).
+  3. extract_all() replays FT chronologically (ELO fully built), then computes
+     features for the NS targets at the end. ELO update + labels are guarded
+     for null scores (NS has no result yet); the 133 model features never use
+     the current score, so they come out identical to training.
+  4. Write ONLY the upcoming rows -> data/upcoming_features.csv
+
+Then: generate_daily_picks.py --features data/upcoming_features.csv --log
+
+Run nightly (heavy: full ELO replay, like training). Read-only on the DB.
+"""
+from __future__ import annotations
+import csv
+import json
+import os
+import sys
+import time
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try:
+        sys.stdout.reconfigure(encoding="utf-8")
+    except Exception:
+        pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.insert(0, AI_DIR)
+
+from scripts.extract_training_data_v27 import (  # noqa: E402
+    V27Loader, V27Extractor, ALL_COLS, get_conn,
+)
+
+OUTPUT = os.path.join(AI_DIR, "data", "upcoming_features.csv")
+DAYS_AHEAD = 4
+
+
+def map_live_odds(odds_json) -> dict:
+    """Map live_matches.odds JSON → odds_cache keys, IDENTICAL to the trainer's
+    _load_odds category/selection logic (so odds features match training)."""
+    out: dict = {}
+    if isinstance(odds_json, str):
+        try:
+            odds_json = json.loads(odds_json)
+        except Exception:
+            return out
+    if not isinstance(odds_json, dict):
+        return out
+    for cat, sels in odds_json.items():
+        if not isinstance(sels, dict):
+            continue
+        c = str(cat).lower().strip()
+        for sel, val in sels.items():
+            try:
+                v = float(val)
+            except (TypeError, ValueError):
+                continue
+            if v <= 0:
+                continue
+            sn = str(sel)
+            s = sn.lower().strip()
+            if c == "maç sonucu":
+                if sn == "1": out["ms_h"] = v
+                elif sn in ("0", "X"): out["ms_d"] = v
+                elif sn == "2": out["ms_a"] = v
+            elif c == "1. yarı sonucu":
+                if sn == "1": out["ht_ms_h"] = v
+                elif sn in ("0", "X"): out["ht_ms_d"] = v
+                elif sn == "2": out["ht_ms_a"] = v
+            elif c == "karşılıklı gol":
+                if "var" in s: out["btts_y"] = v
+                elif "yok" in s: out["btts_n"] = v
+            elif c == "0,5 alt/üst":
+                if "alt" in s: out["ou05_u"] = v
+                elif "üst" in s: out["ou05_o"] = v
+            elif c == "1,5 alt/üst":
+                if "alt" in s: out["ou15_u"] = v
+                elif "üst" in s: out["ou15_o"] = v
+            elif c == "2,5 alt/üst":
+                if "alt" in s: out["ou25_u"] = v
+                elif "üst" in s: out["ou25_o"] = v
+            elif c == "3,5 alt/üst":
+                if "alt" in s: out["ou35_u"] = v
+                elif "üst" in s: out["ou35_o"] = v
+            elif c == "1. yarı 0,5 alt/üst":
+                if "alt" in s: out["ht_ou05_u"] = v
+                elif "üst" in s: out["ht_ou05_o"] = v
+            elif c == "1. yarı 1,5 alt/üst":
+                if "alt" in s: out["ht_ou15_u"] = v
+                elif "üst" in s: out["ht_ou15_o"] = v
+    return out
+
+
+class UpcomingExtractor(V27Extractor):
+    """Same feature computation as training; only guards null (NS) scores."""
+
+    def _update_elo(self, home_id, away_id, score_home, score_away):
+        if score_home is None or score_away is None:
+            return  # upcoming match — no result, don't move ELO
+        return super()._update_elo(home_id, away_id, score_home, score_away)
+
+    def _extract_one(self, mid, hid, aid, sh, sa, hth, hta, mst, lid, hn, an, ln):
+        if sh is None or sa is None:
+            # Upcoming TARGET. Dummy scores so label/total_goals don't crash;
+            # those columns are labels/LEAKY and are NOT among the 133 model
+            # features, so the served feature vector is identical to training.
+            row = super()._extract_one(mid, hid, aid, 0, 0, 0, 0, mst, lid, hn, an, ln)
+            if row:
+                row["_upcoming"] = 1
+            return row
+        # FT match: needed ONLY to advance ELO (extract_all calls _update_elo
+        # afterwards regardless). Skip the expensive per-match feature
+        # computation — that turns a ~6h full extraction into seconds while
+        # producing the IDENTICAL final ELO the upcoming targets read.
+        return None
+
+
+def main():
+    t0 = time.time()
+    conn = get_conn()
+
+    # ── Cheap check FIRST: are there upcoming matches with odds? ──
+    now_ms = int(time.time() * 1000)
+    hi_ms = now_ms + DAYS_AHEAD * 24 * 3600 * 1000
+    cur = conn.cursor()
+    cur.execute(
+        """
+        SELECT lm.id, lm.home_team_id, lm.away_team_id, lm.mst_utc, lm.league_id,
+               ht.name, at.name, l.name, lm.odds
+        FROM live_matches lm
+        JOIN teams ht ON ht.id = lm.home_team_id
+        JOIN teams at ON at.id = lm.away_team_id
+        JOIN leagues l ON l.id = lm.league_id
+        WHERE lm.sport = 'football'
+          AND lm.odds IS NOT NULL
+          AND lm.mst_utc > %s AND lm.mst_utc <= %s
+        ORDER BY lm.mst_utc ASC
+        """,
+        (now_ms, hi_ms),
+    )
+    upcoming = cur.fetchall()
+    targets = []
+    for mid, hid, aid, mst, lid, hn, an, ln, odds_json in upcoming:
+        oc = map_live_odds(odds_json)
+        if "ms_h" not in oc or "ms_a" not in oc:
+            continue  # need MS odds for the policy
+        targets.append((mid, hid, aid, mst, lid, hn, an, ln, oc))
+    print(f"Upcoming NS matches with MS odds (next {DAYS_AHEAD}d): {len(targets)}", flush=True)
+    if not targets:
+        print("⚠️ Nothing to extract. Deploy the 4-day window + let the odds cron\n"
+              "   populate live_matches, then re-run.")
+        conn.close()
+        return
+
+    print("📦 Loading FT history (ELO/form/league/squad caches; heavy) ...", flush=True)
+    loader = V27Loader(conn)
+    loader.load_all()
+    loader.load_league_matches()
+    print(f"  FT matches: {len(loader.matches)}", flush=True)
+
+    for mid, hid, aid, mst, lid, hn, an, ln, oc in targets:
+        loader.odds_cache[mid] = oc
+        loader.matches.append(
+            (mid, hid, aid, None, None, None, None, mst, lid, hn, an, ln)
+        )
+    # NS targets must be processed AFTER all FT (ELO fully built)
+    loader.matches.sort(key=lambda m: m[7] if m[7] is not None else 0)
+    added = len(targets)
+
+    print("🔄 Extracting features (FT replay + upcoming targets) ...", flush=True)
+    ext = UpcomingExtractor(conn, loader)
+    rows = ext.extract_all()
+    up_rows = [r for r in rows if r.get("_upcoming")]
+
+    os.makedirs(os.path.dirname(OUTPUT), exist_ok=True)
+    with open(OUTPUT, "w", newline="", encoding="utf-8") as f:
+        w = csv.DictWriter(f, fieldnames=ALL_COLS, extrasaction="ignore")
+        w.writeheader()
+        w.writerows(up_rows)
+
+    with_odds = sum(1 for r in up_rows if r.get("odds_ms_h", 0) and r["odds_ms_h"] > 0)
+    print(f"\n✅ Wrote {len(up_rows)} upcoming feature rows ({with_odds} with MS odds) → {OUTPUT}")
+    print(f"   Time: {(time.time()-t0)/60:.1f} min")
+    print("   Next: python scripts/generate_daily_picks.py --features data/upcoming_features.csv --log")
+    conn.close()
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/fit_anchor_corrections.py

@@ -0,0 +1,255 @@
+"""Guarded self-correction loop — fits the market-anchor correction table.
+
+What it does (the "tablo üreteci" of the feedback loop):
+
+  1. MEASURE: on settled real-odds matches, per implied-probability band, the
+     gap between the RAW de-vigged probability and the actual rate — for BOTH
+     the home side (ms_home) and the away side (ms_away).
+  2. BRAKE: a band only earns a correction if it passes the safety gates —
+        * min sample  (>= MIN_N matches in the band, fitted on TRAIN window)
+        * shrinkage   (delta = SHRINK x measured gap — never the full gap)
+        * clipping    (|delta| <= CLIP)
+        * materiality (|delta| >= MIN_DELTA, else 0 — don't chase noise)
+  3. PROVE: the candidate table must beat the CURRENTLY ACTIVE corrections
+     out-of-sample (most recent TEST_DAYS, never seen during fitting) on
+     combined home+away ECE. If it doesn't, nothing is written.
+  4. WRITE: versioned artifact `config/market_anchor_corrections.json`
+     (+ timestamped copy under `config/history/`). The engine reads the table
+     at runtime (models/market_anchor.py) — the loop never modifies code.
+
+Run weekly (cron) or manually after big data ingests:
+    python scripts/fit_anchor_corrections.py [--days 540] [--test-days 90]
+    python scripts/fit_anchor_corrections.py --dry-run   # measure only
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import shutil
+import sys
+import time
+from collections import defaultdict
+from typing import Any, Callable, Dict, List, Optional, Tuple
+
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+
+import psycopg2  # noqa: E402
+from psycopg2.extras import RealDictCursor  # noqa: E402
+
+from data.db import get_clean_dsn  # noqa: E402
+from models.market_anchor import (  # noqa: E402
+    away_favorite_delta,
+    home_favorite_delta,
+)
+
+# ── safety gates ─────────────────────────────────────────────────────
+MIN_N = 1500        # band needs this many TRAIN matches to earn a correction
+SHRINK = 0.5        # apply only half of the measured gap
+CLIP = 0.05         # never correct more than 5 points
+MIN_DELTA = 0.004   # below this the correction is noise — emit 0
+ACCEPT_MARGIN = 0.0002  # candidate must beat active combined ECE by this
+
+BANDS: Tuple[Tuple[float, float], ...] = (
+    (0.05, 0.15), (0.15, 0.25), (0.25, 0.35), (0.35, 0.45),
+    (0.45, 0.55), (0.55, 0.65), (0.65, 0.75), (0.75, 0.85), (0.85, 1.01),
+)
+
+REAL_ODDS_MIN_OVERROUND = 0.05
+
+
+def fetch(days: int) -> List[Dict[str, Any]]:
+    since_ms = int((time.time() - days * 86400) * 1000)
+    sql = """
+        SELECT f.implied_home AS p1, f.implied_draw AS px, f.implied_away AS p2,
+               m.mst_utc,
+               (m.winner = 'home')::int AS home_won,
+               (m.winner = 'away')::int AS away_won
+        FROM football_ai_features f
+        JOIN matches m ON m.id = f.match_id
+        WHERE m.sport = 'football'
+          AND m.winner IN ('home', 'away', 'draw')
+          AND f.odds_overround > %s
+          AND m.mst_utc >= %s
+    """
+    out: List[Dict[str, Any]] = []
+    with psycopg2.connect(get_clean_dsn()) as conn:
+        with conn.cursor() as cur:
+            cur.execute("SET statement_timeout = '120s'")
+        with conn.cursor("fit_stream", cursor_factory=RealDictCursor) as cur:
+            cur.itersize = 5000
+            cur.execute(sql, (REAL_ODDS_MIN_OVERROUND, since_ms))
+            for r in cur:
+                p1, px, p2 = r["p1"], r["px"], r["p2"]
+                if p1 is None or px is None or p2 is None:
+                    continue
+                if abs(float(p1) + float(px) + float(p2) - 1.0) > 0.02:
+                    continue
+                out.append({
+                    "p1": float(p1), "p2": float(p2),
+                    "y1": int(r["home_won"]), "y2": int(r["away_won"]),
+                    "mst_utc": int(r["mst_utc"]),
+                })
+    return out
+
+
+def band_of(p: float) -> Optional[int]:
+    for i, (lo, hi) in enumerate(BANDS):
+        if lo <= p < hi:
+            return i
+    return None
+
+
+def fit_candidate(
+    train: List[Dict[str, Any]], pkey: str, ykey: str
+) -> List[Dict[str, Any]]:
+    n = defaultdict(int); sp = defaultdict(float); sy = defaultdict(int)
+    for r in train:
+        b = band_of(r[pkey])
+        if b is None:
+            continue
+        n[b] += 1; sp[b] += r[pkey]; sy[b] += r[ykey]
+    bands: List[Dict[str, Any]] = []
+    for i, (lo, hi) in enumerate(BANDS):
+        if n[i] < MIN_N:
+            continue  # gate: not enough evidence — no correction for this band
+        raw_gap = (sy[i] / n[i]) - (sp[i] / n[i])
+        delta = max(-CLIP, min(CLIP, SHRINK * raw_gap))
+        if abs(delta) < MIN_DELTA:
+            delta = 0.0
+        bands.append({"lo": lo, "hi": hi, "delta": round(delta, 4),
+                      "n": n[i], "raw_gap": round(raw_gap, 4)})
+    return bands
+
+
+def table_delta_fn(table: List[Dict[str, Any]]) -> Callable[[float], float]:
+    def fn(p: float) -> float:
+        for b in table:
+            if b["lo"] <= p < b["hi"]:
+                return b["delta"]
+        return 0.0
+    return fn
+
+
+def ece(rows: List[Dict[str, Any]], pkey: str, ykey: str,
+        delta_fn: Callable[[float], float]) -> float:
+    n = defaultdict(int); sp = defaultdict(float); sy = defaultdict(int)
+    for r in rows:
+        pc = min(0.98, r[pkey] + delta_fn(r[pkey]))
+        b = min(19, int(pc * 20))
+        n[b] += 1; sp[b] += pc; sy[b] += r[ykey]
+    total = sum(n.values())
+    if not total:
+        return 0.0
+    return sum(n[b] * abs(sp[b] / n[b] - sy[b] / n[b]) for b in n) / total
+
+
+def print_bands(title: str, bands: List[Dict[str, Any]]) -> None:
+    print(f"\ncandidate bands — {title} (after gates):")
+    print(f"{'band':>12} {'n':>8} {'raw_gap_pt':>11} {'delta_pt':>9}")
+    for b in bands:
+        print(f"{b['lo']:>5.2f}-{b['hi']:<5.2f} {b['n']:>8} "
+              f"{100 * b['raw_gap']:>11.2f} {100 * b['delta']:>9.2f}")
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--days", type=int, default=540, help="total lookback")
+    ap.add_argument("--test-days", type=int, default=90,
+                    help="most recent window held out for acceptance")
+    ap.add_argument("--dry-run", action="store_true",
+                    help="measure and report only — never write")
+    args = ap.parse_args()
+
+    rows = fetch(args.days)
+    cutoff_ms = int((time.time() - args.test_days * 86400) * 1000)
+    train = [r for r in rows if r["mst_utc"] < cutoff_ms]
+    test = [r for r in rows if r["mst_utc"] >= cutoff_ms]
+    print(f"matches: total={len(rows)}  train={len(train)}  test(OOS)={len(test)}")
+    if len(train) < 10 * MIN_N or len(test) < 2000:
+        print("ABORT: not enough data for a safe fit — keeping active table.")
+        return
+
+    cand_home = fit_candidate(train, "p1", "y1")
+    cand_away = fit_candidate(train, "p2", "y2")
+    print_bands("ms_home", cand_home)
+    print_bands("ms_away", cand_away)
+
+    # active = whatever the engine currently loads (artifact or fallback)
+    ece_h_act = ece(test, "p1", "y1", home_favorite_delta)
+    ece_a_act = ece(test, "p2", "y2", away_favorite_delta)
+    ece_h_cand = ece(test, "p1", "y1", table_delta_fn(cand_home))
+    ece_a_cand = ece(test, "p2", "y2", table_delta_fn(cand_away))
+    ece_h_raw = ece(test, "p1", "y1", lambda _p: 0.0)
+    ece_a_raw = ece(test, "p2", "y2", lambda _p: 0.0)
+
+    print(f"\nOOS ECE (home/away/combined):")
+    print(f"  raw (devig only) : {100 * ece_h_raw:.3f}% / {100 * ece_a_raw:.3f}% "
+          f"/ {100 * (ece_h_raw + ece_a_raw):.3f}%")
+    print(f"  ACTIVE table     : {100 * ece_h_act:.3f}% / {100 * ece_a_act:.3f}% "
+          f"/ {100 * (ece_h_act + ece_a_act):.3f}%")
+    print(f"  CANDIDATE table  : {100 * ece_h_cand:.3f}% / {100 * ece_a_cand:.3f}% "
+          f"/ {100 * (ece_h_cand + ece_a_cand):.3f}%")
+
+    if args.dry_run:
+        print("\n(dry-run: nothing written)")
+        return
+
+    combined_act = ece_h_act + ece_a_act
+    combined_cand = ece_h_cand + ece_a_cand
+    if combined_cand > combined_act - ACCEPT_MARGIN:
+        print("\nREJECTED: candidate does not beat the active table "
+              "out-of-sample. Active corrections stay. (Bu fren tasarim geregi:"
+              " kanitlayamayan tablo yazilmaz.)")
+        return
+
+    cfg_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "config"))
+    path = os.path.join(cfg_dir, "market_anchor_corrections.json")
+    artifact = {
+        "version": time.strftime("%Y-%m-%dT%H:%M:%S"),
+        "fitted_on": {"days": args.days, "test_days": args.test_days,
+                      "n_train": len(train), "n_test": len(test)},
+        "validated": {
+            "ece_home": {"raw": round(ece_h_raw, 5), "active_before": round(ece_h_act, 5),
+                         "candidate_oos": round(ece_h_cand, 5)},
+            "ece_away": {"raw": round(ece_a_raw, 5), "active_before": round(ece_a_act, 5),
+                         "candidate_oos": round(ece_a_cand, 5)},
+        },
+        "gates": {"min_n": MIN_N, "shrink": SHRINK, "clip": CLIP,
+                  "min_delta": MIN_DELTA},
+        "corrections": {"ms_home": cand_home, "ms_away": cand_away},
+    }
+    hist_dir = os.path.join(cfg_dir, "history")
+    os.makedirs(hist_dir, exist_ok=True)
+    if os.path.exists(path):
+        shutil.copy2(path, os.path.join(
+            hist_dir, f"market_anchor_corrections-{int(time.time())}.json"))
+    with open(path, "w", encoding="utf-8") as fh:
+        json.dump(artifact, fh, ensure_ascii=False, indent=2)
+    print(f"\nACCEPTED: wrote {path}")
+
+    # The deployed ai-engine container has NO volume mounts, so the file above
+    # is invisible to it — app_settings is the shared medium. Running engines
+    # re-read it within ~10 minutes (TTL in models/market_anchor.py).
+    try:
+        with psycopg2.connect(get_clean_dsn()) as conn:
+            with conn.cursor() as cur:
+                cur.execute(
+                    """
+                    INSERT INTO app_settings (key, value, updated_at)
+                    VALUES ('market_anchor_corrections', %s, now())
+                    ON CONFLICT (key)
+                    DO UPDATE SET value = EXCLUDED.value, updated_at = now()
+                    """,
+                    (json.dumps(artifact, ensure_ascii=False),),
+                )
+            conn.commit()
+        print("ACCEPTED: upserted app_settings['market_anchor_corrections'] "
+              "(live engines refresh within ~10 min)")
+    except Exception as exc:  # file artifact still written — warn only
+        print(f"WARN: app_settings upsert failed: {exc}")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/generate_daily_picks.py

@@ -0,0 +1,154 @@
+"""
+Generate Daily Picks — the serving picker for the validated favourite policy.
+============================================================================
+Loads the saved leak-free MS model (models/favorite_v1) and applies the
+favourite-band value policy to a set of matches, emitting the day's STAKED
+picks and logging them for forward paper-trade settlement.
+
+Train/serve consistency: features MUST come from the SAME extractor that built
+training_data_v27.csv. Production path = run the extractor nightly INCLUDING
+upcoming (status NS) matches, then point this script at that CSV. Demo path =
+use the tail of the training CSV as stand-in "today" matches (with the real
+result shown, since those are settled).
+
+Policy: bet the MS side with the biggest model_prob - implied edge, ONLY if
+odds in [--lo,--hi] and edge>--margin. Flat 1u. No longshots, no parlays.
+Non-MS markets are NOT staked (efficient -> model error). One bet per match.
+
+Usage:
+  python scripts/generate_daily_picks.py --demo --n 20          # see it work now
+  python scripts/generate_daily_picks.py --features today.csv    # production
+  python scripts/generate_daily_picks.py --settle                # settle paper log
+"""
+from __future__ import annotations
+import argparse, json, os, sys, datetime
+import numpy as np, pandas as pd, xgboost as xgb
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+MODEL_DIR = os.path.join(AI_DIR, "models", "favorite_v1")
+TRAIN_CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+PAPER_LOG = os.path.join(AI_DIR, "data", "paper_trades.csv")
+MS_ODDS = ["odds_ms_h", "odds_ms_d", "odds_ms_a"]
+MS_PICKS = ["1", "X", "2"]
+
+
+def load_model():
+    bst = xgb.Booster(); bst.load_model(os.path.join(MODEL_DIR, "model.json"))
+    with open(os.path.join(MODEL_DIR, "feature_cols.json"), encoding="utf-8") as f:
+        feats = json.load(f)
+    with open(os.path.join(MODEL_DIR, "metadata.json"), encoding="utf-8") as f:
+        meta = json.load(f)
+    return bst, feats, meta
+
+
+def pick_for_rows(df, bst, feats, lo, hi, margin):
+    X = df.reindex(columns=feats).apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    P = bst.predict(xgb.DMatrix(X))                 # [n,3] home/draw/away
+    O = df[MS_ODDS].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    implied = np.where(O > 1.0, 1.0/O, np.nan)
+    edge = np.where(np.isnan(implied), -9.0, P - implied)
+    out = []
+    for i in range(len(df)):
+        k = int(np.argmax(edge[i])); o = float(O[i, k]); e = float(edge[i, k])
+        staked = (e > margin) and (lo <= o < hi)
+        out.append({"idx": i, "pick": MS_PICKS[k], "odds": round(o, 2),
+                    "model_prob": round(float(P[i, k]), 4), "edge": round(e, 4),
+                    "staked": staked})
+    return out
+
+
+def settle():
+    if not os.path.exists(PAPER_LOG):
+        print("No paper_trades.csv yet."); return
+    pt = pd.read_csv(PAPER_LOG)
+    open_bets = pt[pt["result"].isna()] if "result" in pt.columns else pt
+    if open_bets.empty:
+        print("No open bets to settle.");
+    # settle from training CSV scores if present, else needs DB (left as note)
+    src = pd.read_csv(TRAIN_CSV, low_memory=False, usecols=["match_id","score_home","score_away"])
+    sc = src.set_index("match_id")
+    def res(row):
+        if not pd.isna(row.get("result")): return row["result"]
+        m = sc.index == row["match_id"]
+        if not m.any(): return np.nan
+        r = sc[m].iloc[0]; sh, sa = r["score_home"], r["score_away"]
+        if pd.isna(sh): return np.nan
+        outcome = "1" if sh > sa else ("X" if sh == sa else "2")
+        won = (str(row["pick"]) == outcome)
+        return "WON" if won else "LOST"
+    pt["result"] = pt.apply(res, axis=1)
+    pt["pnl"] = pt.apply(lambda r: (r["odds"]-1.0) if r["result"]=="WON"
+                         else (-1.0 if r["result"]=="LOST" else np.nan), axis=1)
+    pt.to_csv(PAPER_LOG, index=False)
+    s = pt.dropna(subset=["pnl"])
+    if len(s):
+        print(f"Settled {len(s)} bets: hit={100*(s['result']=='WON').mean():.1f}%  "
+              f"ROI={100*s['pnl'].sum()/len(s):+.2f}%  net={s['pnl'].sum():+.1f}u")
+    return
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--features", help="CSV of upcoming matches in training schema")
+    ap.add_argument("--demo", action="store_true", help="use tail of training CSV as 'today'")
+    ap.add_argument("--n", type=int, default=20)
+    ap.add_argument("--lo", type=float, default=1.5)
+    ap.add_argument("--hi", type=float, default=2.2)
+    ap.add_argument("--margin", type=float, default=0.03)
+    ap.add_argument("--settle", action="store_true")
+    ap.add_argument("--log", action="store_true", help="append staked picks to paper_trades.csv")
+    args = ap.parse_args()
+
+    if args.settle:
+        settle(); return
+
+    bst, feats, meta = load_model()
+    print(f"Model {meta['version']} (trained {meta['trained_at']}, holdout "
+          f"ROI {meta['holdout_eval']['roi_pct']}%)  band[{args.lo},{args.hi}] margin {args.margin}\n")
+
+    if args.features:
+        df = pd.read_csv(args.features, low_memory=False)
+        demo = False
+    else:
+        df = pd.read_csv(TRAIN_CSV, low_memory=False).sort_values("mst_utc").tail(args.n).reset_index(drop=True)
+        demo = True
+        print("(DEMO: last matches of training CSV as stand-in for today)\n")
+
+    picks = pick_for_rows(df, bst, feats, args.lo, args.hi, args.margin)
+    staked = [p for p in picks if p["staked"]]
+    print(f"{len(df)} matches scanned -> {len(staked)} STAKED MS picks\n")
+    print(f"  {'match_id':<28}{'pick':>5}{'odds':>7}{'model%':>8}{'edge%':>7}" + ("   result" if demo else ""))
+    print("  "+"-"*60)
+    log_rows = []
+    for p in picks:
+        if not p["staked"]: continue
+        r = df.iloc[p["idx"]]; mid = str(r["match_id"])
+        res = ""
+        if demo:
+            sh, sa = r.get("score_home"), r.get("score_away")
+            if pd.notna(sh):
+                out = "1" if sh>sa else ("X" if sh==sa else "2")
+                res = "  WON" if p["pick"]==out else "  lost"
+        print(f"  {mid:<28}{p['pick']:>5}{p['odds']:>7.2f}{100*p['model_prob']:>8.1f}{100*p['edge']:>+7.1f}{res}")
+        log_rows.append({"logged_at": datetime.datetime.now().isoformat(timespec="seconds"),
+                         "match_id": mid, "market": "MS", "pick": p["pick"], "odds": p["odds"],
+                         "model_prob": p["model_prob"], "edge": p["edge"], "stake": 1.0,
+                         "result": np.nan, "pnl": np.nan})
+    if args.log and log_rows and not demo:
+        new = pd.DataFrame(log_rows)
+        if os.path.exists(PAPER_LOG):
+            new = pd.concat([pd.read_csv(PAPER_LOG), new], ignore_index=True)
+        new.to_csv(PAPER_LOG, index=False)
+        print(f"\n  logged {len(log_rows)} picks -> {PAPER_LOG}")
+    elif args.log and demo:
+        print("\n  (--log ignored in --demo; only real upcoming picks are logged)")
+    print("\nReminder: paper-trade only. Stake real money after weeks of forward")
+    print("CLV>0 + ROI>0 (settle with --settle, check scoreboard/clv_report).")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/live_scoreboard.py

@@ -0,0 +1,253 @@
+"""
+Live Scoreboard — the single source of truth for real betting performance.
+=========================================================================
+Reads the *forward-tracked* results in `prediction_runs` (one row per analyzed
+match, with the staked main pick + actual outcome + realized unit_profit) and
+reports what ACTUALLY happened with real money logic — NOT a backtest.
+
+Why this exists: backtests on this codebase are overfit (a paper "+32.7% ROI"
+strategy that the live engine never even ran). The only trustworthy number is
+the realized P/L recorded after matches settle. This tool surfaces it.
+
+Read-only. SELECT only. Safe to run anytime.
+
+Usage:
+  python scripts/live_scoreboard.py
+  python scripts/live_scoreboard.py --days 30
+  python scripts/live_scoreboard.py --version v28-pro-max
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from collections import defaultdict
+from datetime import datetime, timedelta, timezone
+from typing import Any, Dict, List, Optional
+
+# utf-8 stdout so Turkish market/league names never crash on Windows cp1252
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try:
+        sys.stdout.reconfigure(encoding="utf-8")
+    except Exception:
+        pass
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+AI_ENGINE_DIR = os.path.dirname(SCRIPT_DIR)
+sys.path.insert(0, AI_ENGINE_DIR)
+
+from data.db import get_clean_dsn  # noqa: E402
+import psycopg2  # noqa: E402
+from psycopg2.extras import RealDictCursor  # noqa: E402
+
+ODDS_BANDS = [(0, 1.5, "<1.5"), (1.5, 2.0, "1.5-2"), (2.0, 3.0, "2-3"),
+              (3.0, 5.0, "3-5"), (5.0, 6.0, "5-6"), (6.0, 7.5, "6-7.5"),
+              (7.5, 999, "7.5+")]
+
+
+def _f(x: Any, d: Optional[float] = None) -> Optional[float]:
+    try:
+        return float(x) if x is not None else d
+    except (TypeError, ValueError):
+        return d
+
+
+def _parse(j: Any) -> Dict[str, Any]:
+    if isinstance(j, str):
+        try:
+            return json.loads(j)
+        except Exception:
+            return {}
+    return j or {}
+
+
+def _band(odds: Optional[float]) -> str:
+    if odds is None:
+        return "?"
+    for lo, hi, name in ODDS_BANDS:
+        if lo <= odds < hi:
+            return name
+    return "?"
+
+
+def fetch_rows(args) -> List[Dict[str, Any]]:
+    dsn = get_clean_dsn()
+    where = ["eventual_outcome IS NOT NULL"]
+    params: List[Any] = []
+    if args.version:
+        where.append("engine_version = %s")
+        params.append(args.version)
+    if args.days:
+        cutoff = datetime.now(timezone.utc) - timedelta(days=args.days)
+        where.append("generated_at >= %s")
+        params.append(cutoff)
+    sql = f"""
+        SELECT match_id, engine_version, generated_at, eventual_outcome,
+               unit_profit, payload_summary
+        FROM prediction_runs
+        WHERE {' AND '.join(where)}
+        ORDER BY generated_at ASC
+    """
+    with psycopg2.connect(dsn) as conn:
+        with conn.cursor(cursor_factory=RealDictCursor) as cur:
+            cur.execute(sql, params)
+            return cur.fetchall()
+
+
+def distill(rows: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    """One analytic record per run with the staked pick + realized P/L."""
+    out = []
+    for r in rows:
+        ps = _parse(r["payload_summary"])
+        mp = ps.get("main_pick") or {}
+        playable = bool(mp.get("playable"))
+        stake = _f(mp.get("stake_units"), 0.0) or 0.0
+        profit = _f(r["unit_profit"], 0.0) or 0.0
+        outcome = str(r["eventual_outcome"] or "")
+        staked = playable and stake > 0
+        # settled stake = a real bet with a win/loss (exclude NO_BET / push)
+        settled_stake = staked and not outcome.startswith(("NO_BET", "PUSH", "VOID", "CANCEL"))
+        out.append({
+            "match_id": r["match_id"],
+            "version": r["engine_version"],
+            "ts": r["generated_at"],
+            "market": mp.get("market") or "?",
+            "pick": mp.get("pick"),
+            "odds": _f(mp.get("odds")),
+            "stake": stake,
+            "profit": profit,
+            "outcome": outcome,
+            "staked": staked,
+            "settled_stake": settled_stake,
+            "win": settled_stake and profit > 0,
+        })
+    return out
+
+
+def _agg(recs: List[Dict[str, Any]]) -> Dict[str, Any]:
+    # NOTE: recorded unit_profit is on a FLAT 1u basis (win=odds-1, loss=-1),
+    # independent of the brain's suggested stake_units. So ROI is profit per
+    # bet at 1u flat = profit / n. (Using stake_units as denominator is wrong:
+    # it double-counts and produces impossible >100% losses.)
+    s = [r for r in recs if r["settled_stake"]]
+    n = len(s)
+    wins = sum(1 for r in s if r["win"])
+    sug_stake = sum(r["stake"] for r in s)
+    profit = sum(r["profit"] for r in s)
+    return {
+        "n": n,
+        "wins": wins,
+        "hit_pct": round(100.0 * wins / n, 1) if n else None,
+        "sug_stake": round(sug_stake, 2),
+        "profit": round(profit, 2),
+        "roi_pct": round(100.0 * profit / n, 1) if n else None,  # flat 1u
+    }
+
+
+def _line(label: str, a: Dict[str, Any]) -> str:
+    return (f"  {label:<14} n={a['n']:>4}  hit={str(a['hit_pct'] if a['hit_pct'] is not None else '-'):>5}%  "
+            f"profit={a['profit']:>8.2f}u  ROI(flat1u)={str(a['roi_pct'] if a['roi_pct'] is not None else '-'):>7}%")
+
+
+def risk_metrics(recs: List[Dict[str, Any]]) -> Dict[str, Any]:
+    s = [r for r in sorted(recs, key=lambda x: x["ts"]) if r["settled_stake"]]
+    cum = 0.0
+    peak = 0.0
+    max_dd = 0.0
+    streak = 0
+    worst_streak = 0
+    for r in s:
+        cum += r["profit"]
+        peak = max(peak, cum)
+        max_dd = min(max_dd, cum - peak)
+        if r["profit"] <= 0:
+            streak += 1
+            worst_streak = max(worst_streak, streak)
+        else:
+            streak = 0
+    return {"max_drawdown_u": round(max_dd, 2),
+            "longest_losing_streak": worst_streak,
+            "final_cum_u": round(cum, 2)}
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--days", type=int, default=None, help="Only last N days")
+    ap.add_argument("--version", help="Filter by engine_version")
+    args = ap.parse_args()
+
+    rows = fetch_rows(args)
+    recs = distill(rows)
+
+    print("=" * 74)
+    print("LIVE SCOREBOARD  —  realized results from prediction_runs (NOT backtest)")
+    print("=" * 74)
+    if recs:
+        lo = min(r["ts"] for r in recs).date()
+        hi = max(r["ts"] for r in recs).date()
+        print(f"window: {lo} .. {hi}   settled runs: {len(recs)}"
+              + (f"   filter: {args.version}" if args.version else ""))
+    print()
+
+    overall = _agg(recs)
+    print("OVERALL (staked = playable bets only)")
+    print(_line("ALL", overall))
+    no_bet = sum(1 for r in recs if not r["staked"])
+    print(f"  (analyzed {len(recs)} matches; {overall['n']} actually staked, "
+          f"{no_bet} NO_BET)")
+    if overall["n"]:
+        rm = risk_metrics(recs)
+        print(f"  max drawdown: {rm['max_drawdown_u']}u   "
+              f"longest losing streak: {rm['longest_losing_streak']}   "
+              f"net: {rm['final_cum_u']}u")
+    print()
+
+    print("BY ENGINE VERSION")
+    by_v = defaultdict(list)
+    for r in recs:
+        by_v[r["version"]].append(r)
+    for v, rs in sorted(by_v.items(), key=lambda kv: -len(kv[1])):
+        print(_line(v, _agg(rs)))
+    print()
+
+    print("BY MARKET (staked)")
+    by_m = defaultdict(list)
+    for r in recs:
+        if r["settled_stake"]:
+            by_m[r["market"]].append(r)
+    for m, rs in sorted(by_m.items(), key=lambda kv: -len(kv[1])):
+        print(_line(m, _agg(rs)))
+    if not by_m:
+        print("  (no staked settled bets in window)")
+    print()
+
+    print("BY ODDS BAND (staked)")
+    by_b = defaultdict(list)
+    for r in recs:
+        if r["settled_stake"]:
+            by_b[_band(r["odds"])].append(r)
+    for _, _, name in ODDS_BANDS:
+        if name in by_b:
+            print(_line(name, _agg(by_b[name])))
+    print()
+
+    print("WEEKLY TREND (staked)")
+    by_w = defaultdict(list)
+    for r in recs:
+        if r["settled_stake"]:
+            iso = r["ts"].isocalendar()
+            by_w[f"{iso[0]}-W{iso[1]:02d}"].append(r)
+    for w in sorted(by_w):
+        a = _agg(by_w[w])
+        print(_line(w, a))
+    print()
+    print("=" * 74)
+    print("READ: ROI < 0 over a meaningful sample = the staked signals are not")
+    print("profitable. 'NO_BET' rows are free (no stake). CLV is unmeasurable")
+    print("until odds movement is captured (see scripts + odds_history fix).")
+    print("=" * 74)
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/market_calibration.py

@@ -0,0 +1,162 @@
+"""
+Market Calibration Scan — find where the ODDS THEMSELVES are systematically wrong.
+=================================================================================
+The legit, measurable version of "odds şike": pockets (leagues / teams / bands)
+where the market's implied probability does NOT match realized frequency, so a
+SIMPLE rule (no model) is +EV. This is pure market inefficiency — soft pricing
+in obscure leagues, persistent team bias, etc.
+
+Discipline against false 'rigged' pockets (the multiple-comparison trap):
+  * split history by time into HALF-1 (discover) and HALF-2 (validate)
+  * a pocket counts ONLY if it is +EV in BOTH halves with enough bets each
+  * report realized-vs-implied gap (the miscalibration) + ROI
+
+No model. Just odds vs outcomes. Read-only on the training CSV (104k matches
+with odds). Forward 'suspicious line movement' detection needs odds_history
+(currently empty) — separate, forward-only.
+
+Usage: python scripts/market_calibration.py --min-bets 120 --side fav
+"""
+from __future__ import annotations
+import argparse, os, sys
+import numpy as np, pandas as pd
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+
+
+def league_names(ids):
+    try:
+        sys.path.insert(0, AI_DIR)
+        from data.db import get_clean_dsn
+        import psycopg2
+        from psycopg2.extras import RealDictCursor
+        ids = [str(i) for i in ids if i is not None]
+        for _ in range(3):
+            try:
+                with psycopg2.connect(get_clean_dsn()) as c:
+                    with c.cursor(cursor_factory=RealDictCursor) as cur:
+                        cur.execute("SELECT id,name FROM leagues WHERE id = ANY(%s)", (ids,))
+                        return {str(r["id"]): r["name"] for r in cur.fetchall()}
+            except Exception:
+                import time; time.sleep(1)
+    except Exception:
+        pass
+    return {}
+
+
+def team_names(ids):
+    try:
+        sys.path.insert(0, AI_DIR)
+        from data.db import get_clean_dsn
+        import psycopg2
+        from psycopg2.extras import RealDictCursor
+        ids = [str(i) for i in ids if i is not None]
+        for _ in range(3):
+            try:
+                with psycopg2.connect(get_clean_dsn()) as c:
+                    with c.cursor(cursor_factory=RealDictCursor) as cur:
+                        cur.execute("SELECT id,name FROM teams WHERE id = ANY(%s)", (ids,))
+                        return {str(r["id"]): r["name"] for r in cur.fetchall()}
+            except Exception:
+                import time; time.sleep(1)
+    except Exception:
+        pass
+    return {}
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--min-bets", type=int, default=120, help="min bets PER HALF")
+    ap.add_argument("--fav-max", type=float, default=2.5, help="only count favourites below this odds")
+    args = ap.parse_args()
+
+    df = pd.read_csv(CSV, low_memory=False,
+                     usecols=["match_id","league_id","home_team_id","away_team_id","mst_utc",
+                              "odds_ms_h","odds_ms_d","odds_ms_a","score_home","score_away"])
+    df = df.sort_values("mst_utc").reset_index(drop=True)
+    sh = pd.to_numeric(df["score_home"],errors="coerce"); sa = pd.to_numeric(df["score_away"],errors="coerce")
+    ok = sh.notna()&sa.notna()
+    df = df[ok].reset_index(drop=True); sh=sh[ok.values].values; sa=sa[ok.values].values
+    O = df[["odds_ms_h","odds_ms_d","odds_ms_a"]].apply(pd.to_numeric,errors="coerce").fillna(0.0).values
+    valid = (O>1.0).all(1)
+    outcome = np.where(sh>sa,0,np.where(sh==sa,1,2))   # 0 home,1 draw,2 away
+    fav = O.argmin(1); fav_odds = O[np.arange(len(O)),fav]
+    fav_won = (fav==outcome).astype(float)
+    fav_implied = 1.0/fav_odds
+    pnl = np.where(fav_won, fav_odds-1.0, -1.0)
+    half = (np.arange(len(df)) >= len(df)//2).astype(int)   # 0=first half,1=second
+    use = valid & (fav_odds <= args.fav_max)
+
+    base = pd.DataFrame({
+        "league": df["league_id"].astype(str).values,
+        "home": df["home_team_id"].astype(str).values,
+        "fav_is_home": (fav==0),
+        "won": fav_won, "implied": fav_implied, "pnl": pnl, "half": half, "use": use,
+        "fav_odds": fav_odds,
+    })
+    b = base[base["use"]].copy()
+    print(f"{len(b):,} favourite bets (odds<= {args.fav_max}); split into 2 time halves\n")
+    print(f"GLOBAL favourite: realized={100*b['won'].mean():.1f}% implied={100*b['implied'].mean():.1f}% "
+          f"ROI={100*b['pnl'].mean():+.2f}%  (negative = vig; market roughly right)")
+
+    def scan(groupcol, label, namefn, min_bets):
+        rows=[]
+        for key,d in b.groupby(groupcol):
+            h0=d[d["half"]==0]; h1=d[d["half"]==1]
+            if len(h0)<min_bets or len(h1)<min_bets: continue
+            r0=100*h0["pnl"].mean(); r1=100*h1["pnl"].mean()
+            # miscalibration gap: realized - implied (positive = market underprices the favourite)
+            gap=100*(d["won"].mean()-d["implied"].mean())
+            both_pos = r0>0 and r1>0
+            rows.append((min(r0,r1), key, len(d), 100*d["pnl"].mean(), r0, r1, gap, both_pos))
+        rows.sort(reverse=True)
+        names = namefn([r[1] for r in rows[:40]])
+        print(f"\n{'='*82}\n{label}  (✓ = +EV in BOTH halves, the only trustworthy ones)\n{'='*82}")
+        print(f"  {'name':<30}{'n':>6}{'ROI%':>7}{'H1%':>7}{'H2%':>7}{'gap%':>7}  ✓")
+        print("  "+"-"*72)
+        shown=0
+        for mn,key,n,roi,r0,r1,gap,both in rows:
+            if shown>=20 and not both: continue
+            nm=(names.get(key,key) or key)[:28]
+            mark = "✓" if both else ""
+            print(f"  {nm:<30}{n:>6}{roi:>+7.1f}{r0:>+7.1f}{r1:>+7.1f}{gap:>+7.1f}  {mark}")
+            shown+=1
+            if shown>=25: break
+        good=[r for r in rows if r[7]]
+        print(f"\n  -> {len(good)} {label.split()[0].lower()} pockets are +EV in BOTH halves "
+              f"(out of {len(rows)} with enough data)")
+        return good
+
+    scan("league", "BY LEAGUE (favourite flat bet)", league_names, args.min_bets)
+    # team: only when the team is the home favourite (cleanest, most samples)
+    bt = b[b["fav_is_home"]]
+    globals()['b'] = bt  # reuse scan on home-favourite subset
+    # inline team scan
+    rows=[]
+    for key,d in bt.groupby("home"):
+        h0=d[d["half"]==0]; h1=d[d["half"]==1]
+        if len(h0)<max(25,args.min_bets//3) or len(h1)<max(25,args.min_bets//3): continue
+        r0=100*h0["pnl"].mean(); r1=100*h1["pnl"].mean()
+        gap=100*(d["won"].mean()-d["implied"].mean())
+        rows.append((min(r0,r1), key, len(d), 100*d["pnl"].mean(), r0, r1, gap, r0>0 and r1>0))
+    rows.sort(reverse=True)
+    tn=team_names([r[1] for r in rows[:40]])
+    print(f"\n{'='*82}\nBY TEAM as HOME FAVOURITE  (✓ = +EV both halves)\n{'='*82}")
+    print(f"  {'team':<30}{'n':>6}{'ROI%':>7}{'H1%':>7}{'H2%':>7}{'gap%':>7}  ✓")
+    print("  "+"-"*72)
+    for mn,key,n,roi,r0,r1,gap,both in rows[:22]:
+        nm=(tn.get(key,key) or key)[:28]; mark="✓" if both else ""
+        print(f"  {nm:<30}{n:>6}{roi:>+7.1f}{r0:>+7.1f}{r1:>+7.1f}{gap:>+7.1f}  {mark}")
+    good=[r for r in rows if r[7]]
+    print(f"\n  -> {len(good)} teams +EV in BOTH halves (out of {len(rows)})")
+    print("\nREAD: ✓ pockets survived a time-split = candidate real inefficiencies (not noise).")
+    print("Still forward-validate with CLV. No ✓ = market is efficient there; don't bet.")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/match_report.py

@@ -0,0 +1,151 @@
+"""
+Match Report — calibrated outcome probabilities + loss-minimizing pick per match.
+================================================================================
+For each match, shows the model's CALIBRATED probability for every outcome
+(1X2, Double Chance, OU 1.5/2.5/3.5, BTTS, HT), next to the market's implied
+probability, and recommends:
+  * EN GÜVENLİ  = highest-probability outcome (most likely to hit / lowest variance)
+  * EN İYİ DEĞER = least-negative-EV outcome (smartest bet given the margin)
+
+Probabilities are leak-free and calibrated (ECE ~0.43%, see calibration_report).
+This is a LOSS-MINIMIZER, not a profit machine — accurate probabilities to make
+the smartest, least-losing decisions against İddaa's high margin.
+
+Trains the market models on the full history (leak-free), then scores the input.
+
+Usage:
+  python scripts/match_report.py --features data/upcoming_features.csv
+  python scripts/match_report.py --demo --n 6
+"""
+from __future__ import annotations
+import argparse, os, sys, time
+import numpy as np, pandas as pd, xgboost as xgb
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.insert(0, AI_DIR)
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+META = {"match_id","home_team_id","away_team_id","league_id","mst_utc",
+        "score_home","score_away","ht_score_home","ht_score_away"}
+LEAKY = {"home_goals_form","away_goals_form","total_goals","ht_total_goals",
+         "squad_diff","home_squad_quality","away_squad_quality",
+         "referee_home_bias","referee_avg_goals"}
+
+def ou(line):  return lambda sh,sa,hh,ha: (0 if (sh+sa) > line else 1)
+def htou(line):return lambda sh,sa,hh,ha: (None if np.isnan(hh) else (0 if (hh+ha)>line else 1))
+MARKETS = {
+  "MS":   ("multi", ["odds_ms_h","odds_ms_d","odds_ms_a"], ["1","X","2"],
+           lambda sh,sa,hh,ha: 0 if sh>sa else (1 if sh==sa else 2)),
+  "OU15": ("binary",["odds_ou15_o","odds_ou15_u"], ["1.5 Üst","1.5 Alt"], ou(1.5)),
+  "OU25": ("binary",["odds_ou25_o","odds_ou25_u"], ["2.5 Üst","2.5 Alt"], ou(2.5)),
+  "OU35": ("binary",["odds_ou35_o","odds_ou35_u"], ["3.5 Üst","3.5 Alt"], ou(3.5)),
+  "BTTS": ("binary",["odds_btts_y","odds_btts_n"], ["KG Var","KG Yok"],
+           lambda sh,sa,hh,ha: 0 if (sh>0 and sa>0) else 1),
+  "HT":   ("multi", ["odds_ht_ms_h","odds_ht_ms_d","odds_ht_ms_a"], ["İY 1","İY X","İY 2"],
+           lambda sh,sa,hh,ha: None if np.isnan(hh) else (0 if hh>ha else (1 if hh==ha else 2))),
+}
+PM={"objective":"multi:softprob","num_class":3,"max_depth":5,"eta":0.05,"subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+PB={"objective":"binary:logistic","max_depth":5,"eta":0.05,"subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+
+
+def team_names(ids):
+    try:
+        from data.db import get_clean_dsn
+        import psycopg2; from psycopg2.extras import RealDictCursor
+        ids=[str(i) for i in ids]
+        for _ in range(3):
+            try:
+                with psycopg2.connect(get_clean_dsn()) as c:
+                    with c.cursor(cursor_factory=RealDictCursor) as cur:
+                        cur.execute("SELECT id,name FROM teams WHERE id = ANY(%s)",(ids,))
+                        return {str(r["id"]):r["name"] for r in cur.fetchall()}
+            except Exception: time.sleep(1)
+    except Exception: pass
+    return {}
+
+
+def main():
+    ap=argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--features"); ap.add_argument("--demo",action="store_true")
+    ap.add_argument("--n",type=int,default=8); ap.add_argument("--estimators",type=int,default=250)
+    args=ap.parse_args()
+
+    df=pd.read_csv(CSV,low_memory=False).sort_values("mst_utc").reset_index(drop=True)
+    sh=pd.to_numeric(df["score_home"],errors="coerce"); sa=pd.to_numeric(df["score_away"],errors="coerce")
+    ok=sh.notna()&sa.notna(); df=df[ok].reset_index(drop=True)
+    SH=sh[ok.values].values.astype(float); SA=sa[ok.values].values.astype(float)
+    HH=pd.to_numeric(df["ht_score_home"],errors="coerce").values.astype(float)
+    HA=pd.to_numeric(df["ht_score_away"],errors="coerce").values.astype(float)
+    feats=[c for c in df.columns if c not in META and not c.startswith("label_") and c not in LEAKY]
+    X=df[feats].apply(pd.to_numeric,errors="coerce").fillna(0.0).values
+    N=len(df)
+    print(f"Training {len(MARKETS)} leak-free calibrated market models on {N:,} matches ...",flush=True)
+    models={}
+    for m,(kind,oc,picks,tfn) in MARKETS.items():
+        truth=np.array([tfn(SH[i],SA[i],HH[i],HA[i]) for i in range(N)],dtype=object)
+        valid=np.array([v is not None for v in truth])
+        if kind=="multi":
+            b=xgb.train(PM,xgb.DMatrix(X[valid],label=truth[valid].astype(int)),num_boost_round=args.estimators)
+        else:
+            b=xgb.train(PB,xgb.DMatrix(X[valid],label=(truth[valid].astype(int)==0).astype(int)),num_boost_round=args.estimators)
+        models[m]=(kind,oc,picks,b)
+
+    # input matches
+    if args.features:
+        inp=pd.read_csv(args.features,low_memory=False); demo=False
+    else:
+        inp=df.tail(args.n).reset_index(drop=True); demo=True
+        print("(DEMO: training CSV son maçları)\n")
+    names=team_names(list(inp.get("home_team_id",[]))+list(inp.get("away_team_id",[]))) if "home_team_id" in inp.columns else {}
+    Xi=inp.reindex(columns=feats).apply(pd.to_numeric,errors="coerce").fillna(0.0).values
+
+    shown=0
+    for i in range(len(inp)):
+        if shown>=args.n: break
+        r=inp.iloc[i]; xrow=Xi[i:i+1]
+        hn=names.get(str(r.get("home_team_id")),str(r.get("home_team_id","?"))[:8])
+        an=names.get(str(r.get("away_team_id")),str(r.get("away_team_id","?"))[:8])
+        print("="*68)
+        print(f"{hn} vs {an}")
+        print(f"  {'market':<8}{'sonuç':<10}{'model%':>8}{'piyasa%':>9}{'oran':>7}{'EV%':>8}")
+        print("  "+"-"*58)
+        bets=[]; ms_probs=None
+        for m,(kind,oc,picks,b) in models.items():
+            if kind=="multi":
+                P=b.predict(xgb.DMatrix(xrow))[0]
+            else:
+                p=float(b.predict(xgb.DMatrix(xrow))[0]); P=np.array([p,1-p])
+            if m=="MS": ms_probs=P
+            O=pd.to_numeric(r.reindex(oc),errors="coerce").fillna(0.0).values
+            for k in range(len(picks)):
+                o=float(O[k]); mp=float(P[k])
+                if o>1.0:
+                    imp=1/o; ev=mp*o-1
+                    print(f"  {m:<8}{picks[k]:<10}{100*mp:>7.0f}%{100*imp:>8.0f}%{o:>7.2f}{100*ev:>+7.1f}")
+                    bets.append((m,picks[k],mp,o,ev))
+                else:
+                    print(f"  {m:<8}{picks[k]:<10}{100*mp:>7.0f}%{'-':>8} {'-':>6} {'-':>7}")
+        # Double Chance derived from MS (no odds shown — Nesine'de oranına bakarsın)
+        if ms_probs is not None:
+            h,d,a=ms_probs
+            print(f"  {'DC':<8}{'1X':<10}{100*(h+d):>7.0f}%   (türetilmiş 'en güvenli' seçenek)")
+            print(f"  {'DC':<8}{'X2':<10}{100*(d+a):>7.0f}%")
+            print(f"  {'DC':<8}{'12':<10}{100*(h+a):>7.0f}%")
+        print("  "+"-"*58)
+        if bets:
+            safe=max(bets,key=lambda x:x[2])     # highest probability
+            value=max(bets,key=lambda x:x[4])    # least-negative EV
+            print(f"  >>> EN GÜVENLİ : {safe[0]} {safe[1]} (model %{100*safe[2]:.0f}, oran {safe[3]:.2f})")
+            print(f"  >>> EN İYİ DEĞER: {value[0]} {value[1]} (EV %{100*value[4]:+.1f}, model %{100*value[2]:.0f}, oran {value[3]:.2f})")
+            if value[4] <= 0:
+                print(f"      (EV negatif → marj yüzünden 'kâr' yok; en az kaybettiren bu. Değer yoksa PAS geç.)")
+        shown+=1
+    print("\nNOT: olasılıklar kalibre (model %X ⇒ gerçekte ~%X). EV<0 her yerde olabilir")
+    print("(İddaa marjı); amaç KAYBI MİNİMİZE etmek + en doğru maç okumasını görmek.")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/monitor_odds_movement.py

@@ -0,0 +1,134 @@
+"""
+Odds Movement Monitor — opening→closing line movement + steam radar.
+===================================================================
+Reads live_odds_history (filled by data-fetcher.task.ts every 15 min for
+upcoming matches, all markets) and reports, PER MATCH:
+  * opening odd (first capture) vs closing odd (latest capture)
+  * total move % = (closing - opening) / opening   ← the headline signal
+  * the steam side (the selection that shortened the most = money/info/şike)
+
+Why opening→closing matters: it is the market's TOTAL revision. A side that
+shortened a lot from open to close = the market learned something. If you can
+bet EARLY (before the shortening), that gap is real value (positive CLV) — the
+one realistic edge vs İddaa. As a closing bettor it's a RISK FILTER: heavy
+late steam against your pick = skip.
+
+Capture is done by the NestJS cron now (DB); this is a pure READER.
+
+Usage:
+  python scripts/monitor_odds_movement.py                 # MS movers
+  python scripts/monitor_odds_movement.py --min-move 0.08 --market "Maç Sonucu"
+"""
+from __future__ import annotations
+import argparse, os, sys, time
+from collections import defaultdict
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.insert(0, AI_DIR)
+from data.db import get_clean_dsn  # noqa: E402
+import psycopg2  # noqa: E402
+from psycopg2.extras import RealDictCursor  # noqa: E402
+
+
+def connect():
+    last = None
+    for _ in range(8):
+        try:
+            return psycopg2.connect(get_clean_dsn())
+        except Exception as e:
+            last = e; time.sleep(3)
+    raise last
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--min-move", type=float, default=0.05,
+                    help="flag matches whose focus-market move >= this fraction (default 0.05)")
+    ap.add_argument("--market", default="Maç Sonucu", help="focus market for the watchlist")
+    ap.add_argument("--limit", type=int, default=25)
+    args = ap.parse_args()
+
+    with connect() as c, c.cursor(cursor_factory=RealDictCursor) as cur:
+        cur.execute("SELECT to_regclass('public.live_odds_history') AS ex")
+        if not cur.fetchall()[0]["ex"]:
+            print("live_odds_history yok — NestJS cron'u henüz yazmamış (deploy/build kontrol)."); return
+
+        # opening (earliest) + closing (latest) per match/market/selection
+        cur.execute("""
+            SELECT match_id, market, selection,
+                   (array_agg(new_value ORDER BY change_time ASC))[1]  AS opening,
+                   (array_agg(new_value ORDER BY change_time DESC))[1] AS closing,
+                   count(*) AS ticks
+            FROM live_odds_history
+            GROUP BY match_id, market, selection
+        """)
+        rows = cur.fetchall()
+        if not rows:
+            print("live_odds_history boş (henüz yakalama yok)."); return
+
+        # per match aggregation
+        by_match = defaultdict(lambda: {"focus": {}, "any_ticks": 0, "max_abs": 0.0})
+        for r in rows:
+            mid = r["match_id"]; o = r["opening"]; cl = r["closing"]
+            d = by_match[mid]
+            d["any_ticks"] = max(d["any_ticks"], r["ticks"])
+            if o and cl and o > 0:
+                mv = (cl - o) / o
+                d["max_abs"] = max(d["max_abs"], abs(mv))
+                if r["market"] == args.market:
+                    d["focus"][r["selection"]] = (o, cl, mv)
+
+        # team names + kickoff
+        ids = list(by_match.keys())
+        names = {}
+        if ids:
+            cur.execute("""SELECT lm.id, ht.name h, at.name a, lm.mst_utc
+                           FROM live_matches lm
+                           JOIN teams ht ON ht.id=lm.home_team_id
+                           JOIN teams at ON at.id=lm.away_team_id
+                           WHERE lm.id = ANY(%s)""", (ids,))
+            for r in cur.fetchall():
+                names[r["id"]] = (f"{r['h']} v {r['a']}", r["mst_utc"])
+
+    moved = [(m, d) for m, d in by_match.items() if d["any_ticks"] > 1]
+    print("="*78)
+    print("ODDS MOVEMENT — açılış→kapanış (live_odds_history)")
+    print("="*78)
+    print(f"izlenen maç: {len(by_match)} | hareket başlamış (>1 yakalama): {len(moved)}")
+    if not moved:
+        print("\nHenüz hareket yok — hepsi tek yakalama (açılış). Oranlar oynadıkça dolacak.")
+        print("(NestJS 15-dk cron'u her tazelemede değişen oranı ekliyor.)")
+        return
+
+    flagged = sorted(
+        [(m, d) for m, d in moved if d["focus"] and d["max_abs"] >= args.min_move],
+        key=lambda x: -x[1]["max_abs"],
+    )
+    now = int(time.time()*1000)
+    print(f"\n{args.market} hareketi >= %{args.min_move*100:.0f} olan maçlar:")
+    print(f"  {'maç':<32}{'sel':>5}{'açılış':>8}{'kapanış':>9}{'hareket':>9}")
+    print("  "+"-"*64)
+    for mid, d in flagged[:args.limit]:
+        nm, mst = names.get(mid, (mid[:30], None))
+        ko = ""
+        if mst:
+            mins = (mst-now)/60000
+            ko = f"  KO~{mins/60:.1f}h" if mins > 0 else "  (başladı)"
+        # steam side = most shortened (most negative move)
+        steam = min(d["focus"].items(), key=lambda kv: kv[1][2])
+        print(f"  {nm[:30]:<32}{'':>5}{'':>8}{'':>9}{'':>9}{ko}")
+        for sel, (o, cl, mv) in d["focus"].items():
+            tag = " ↓STEAM" if sel == steam[0] and mv < 0 else ""
+            print(f"  {'':<32}{sel:>5}{o:>8.2f}{cl:>9.2f}{100*mv:>+8.1f}%{tag}")
+    if not flagged:
+        print("  (eşiği geçen yok — hareketler küçük)")
+    print("\nOKUMA: kapanışta oynuyorsan, pick'ine KARŞI ↓STEAM olan maçı PAS geç.")
+    print("Erken oynayabiliyorsan, kısalan tarafı açılışta yakalamak = gerçek değer (CLV).")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/multi_market_edge.py

@@ -0,0 +1,182 @@
+"""
+Multi-Market Edge + Best-Bet Selector — pick the best value bet PER MATCH
+========================================================================
+Not "play the handed main_pick". For each match, score EVERY market the model
+covers, compare model prob vs market implied, and select the single best VALUE
+bet across all markets. Leak-free, walk-forward, honest.
+
+Markets (truth derived from scores, not trusted labels):
+  MS(1X2), HT-result, OU0.5/1.5/2.5/3.5, HT_OU0.5/1.5, BTTS.
+
+Outputs:
+  (A) per-market value ROI  -> which bet types actually carry edge
+  (B) cross-market SELECTOR -> best value bet per match, with odds-band filter,
+      fold-consistency, and the model-free baseline.
+
+⚠️ CSV odds are a static capture, not verified closing. Positive = LEAD; forward
+paper-trade with real CLV before staking.
+
+Usage: python scripts/multi_market_edge.py --folds 5 --lo 1.5 --hi 2.6 --margin 0.03
+"""
+from __future__ import annotations
+import argparse, os, sys
+import numpy as np, pandas as pd, xgboost as xgb
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+META = {"match_id","home_team_id","away_team_id","league_id","mst_utc",
+        "score_home","score_away","ht_score_home","ht_score_away"}
+LEAKY = {"home_goals_form","away_goals_form","total_goals","ht_total_goals",
+         "squad_diff","home_squad_quality","away_squad_quality",
+         "referee_home_bias","referee_avg_goals"}
+
+# market -> (kind, [odds_cols aligned to classes], truth_fn(sh,sa,hh,ha)->class idx or None)
+def ou(line):   return lambda sh,sa,hh,ha: (0 if (sh+sa) > line else 1)            # 0=Over,1=Under
+def htou(line): return lambda sh,sa,hh,ha: (None if np.isnan(hh) else (0 if (hh+ha) > line else 1))
+def ms_truth(sh,sa,hh,ha): return 0 if sh>sa else (1 if sh==sa else 2)
+def ht_truth(sh,sa,hh,ha): return None if np.isnan(hh) else (0 if hh>ha else (1 if hh==ha else 2))
+def btts_truth(sh,sa,hh,ha): return 0 if (sh>0 and sa>0) else 1                    # 0=Yes,1=No
+
+MARKETS = {
+  "MS":      ("multi",  ["odds_ms_h","odds_ms_d","odds_ms_a"], ["1","X","2"], ms_truth),
+  "HT":      ("multi",  ["odds_ht_ms_h","odds_ht_ms_d","odds_ht_ms_a"], ["1","X","2"], ht_truth),
+  "OU05":    ("binary", ["odds_ou05_o","odds_ou05_u"], ["Üst","Alt"], ou(0.5)),
+  "OU15":    ("binary", ["odds_ou15_o","odds_ou15_u"], ["Üst","Alt"], ou(1.5)),
+  "OU25":    ("binary", ["odds_ou25_o","odds_ou25_u"], ["Üst","Alt"], ou(2.5)),
+  "OU35":    ("binary", ["odds_ou35_o","odds_ou35_u"], ["Üst","Alt"], ou(3.5)),
+  "HT_OU05": ("binary", ["odds_ht_ou05_o","odds_ht_ou05_u"], ["Üst","Alt"], htou(0.5)),
+  "HT_OU15": ("binary", ["odds_ht_ou15_o","odds_ht_ou15_u"], ["Üst","Alt"], htou(1.5)),
+  "BTTS":    ("binary", ["odds_btts_y","odds_btts_n"], ["Var","Yok"], btts_truth),
+}
+PARAMS_M = {"objective":"multi:softprob","num_class":3,"max_depth":5,"eta":0.05,
+            "subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+PARAMS_B = {"objective":"binary:logistic","max_depth":5,"eta":0.05,
+            "subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--folds", type=int, default=5)
+    ap.add_argument("--estimators", type=int, default=150)
+    ap.add_argument("--lo", type=float, default=1.5)
+    ap.add_argument("--hi", type=float, default=2.6)
+    ap.add_argument("--margin", type=float, default=0.03)
+    args = ap.parse_args()
+
+    df = pd.read_csv(CSV, low_memory=False).sort_values("mst_utc").reset_index(drop=True)
+    sh = pd.to_numeric(df["score_home"], errors="coerce")
+    sa = pd.to_numeric(df["score_away"], errors="coerce")
+    ok = sh.notna() & sa.notna()
+    df = df[ok].reset_index(drop=True)
+    SH = sh[ok.values].values.astype(float); SA = sa[ok.values].values.astype(float)
+    HH = pd.to_numeric(df["ht_score_home"], errors="coerce").values.astype(float)
+    HA = pd.to_numeric(df["ht_score_away"], errors="coerce").values.astype(float)
+    feats = [c for c in df.columns if c not in META and not c.startswith("label_") and c not in LEAKY]
+    X = df[feats].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    N = len(df)
+    print(f"{N:,} matches, {len(feats)} leak-free feats, {len(MARKETS)} markets, folds={args.folds}")
+
+    # precompute truth + odds per market
+    MK = {}
+    for mname,(kind,ocols,picks,tfn) in MARKETS.items():
+        if not all(c in df.columns for c in ocols):
+            print(f"  skip {mname}: missing odds cols"); continue
+        O = df[ocols].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+        truth = np.array([tfn(SH[i],SA[i],HH[i],HA[i]) for i in range(N)], dtype=object)
+        MK[mname] = (kind, O, picks, truth)
+
+    start = int(N*0.5); bounds = np.linspace(start, N, args.folds+1, dtype=int)
+
+    # accumulators
+    per_market = {m: {"n":0,"pnl":0.0,"win":0} for m in MK}                 # (A) best value pick within market
+    sel = {"n":0,"pnl":0.0,"win":0,"fold":{}}                                # (B) cross-market selector
+    sel_by_mkt = {m: {"n":0,"pnl":0.0,"win":0} for m in MK}
+
+    for fi in range(args.folds):
+        te0,te1 = bounds[fi], bounds[fi+1]
+        if te1-te0 < 50: continue
+        idx = np.arange(te0,te1)
+        # train each market model on [:te0], predict test
+        cand = {}  # market -> (P_matrix[n_test, n_picks], O_test, truth_test)
+        for m,(kind,O,picks,truth) in MK.items():
+            ytr_full = truth[:te0]
+            # mask invalid truth (e.g., HT markets with missing HT score)
+            valid_tr = np.array([v is not None for v in ytr_full])
+            if kind=="multi":
+                ytr = ytr_full[valid_tr].astype(int)
+                bst = xgb.train(PARAMS_M, xgb.DMatrix(X[:te0][valid_tr], label=ytr), num_boost_round=args.estimators)
+                P = bst.predict(xgb.DMatrix(X[te0:te1]))           # [n,3]
+            else:
+                ytr = ytr_full[valid_tr].astype(int)               # 0=positive,1=neg
+                pos = (ytr==0).astype(int)
+                bst = xgb.train(PARAMS_B, xgb.DMatrix(X[:te0][valid_tr], label=pos), num_boost_round=args.estimators)
+                ppos = bst.predict(xgb.DMatrix(X[te0:te1]))
+                P = np.column_stack([ppos, 1.0-ppos])              # [n,2] -> [pos,neg]
+            cand[m] = (P, O[te0:te1], truth[te0:te1])
+
+        # iterate test matches
+        for j in range(te1-te0):
+            best = None  # (edge, market, pickidx, odds, won)
+            for m,(P,Ot,Tt) in cand.items():
+                t = Tt[j]
+                if t is None: continue
+                probs = P[j]; odds = Ot[j]
+                for k in range(len(probs)):
+                    o = odds[k]
+                    if o <= 1.0: continue
+                    edge = probs[k] - 1.0/o
+                    won = int(t==k)
+                    # (A) per-market: track best value pick in this market (any band, edge>margin)
+                    if edge > args.margin:
+                        d = per_market[m]
+                        # only count the market's single best pick per match
+                    # collect for selector if in band + margin
+                    if edge > args.margin and args.lo <= o < args.hi:
+                        if best is None or edge > best[0]:
+                            best = (edge, m, k, o, won)
+                # per-market best pick (separate loop for clean per-market ROI in band)
+                bestk=None
+                for k in range(len(probs)):
+                    o=odds[k]
+                    if o<=1.0: continue
+                    e=probs[k]-1.0/o
+                    if e>args.margin and args.lo<=o<args.hi and (bestk is None or e>bestk[0]):
+                        bestk=(e,k,o,int(t==k))
+                if bestk is not None:
+                    e,k,o,won = bestk
+                    pnl = (o-1.0) if won else -1.0
+                    d=per_market[m]; d["n"]+=1; d["pnl"]+=pnl; d["win"]+=won
+            # selector: single best value bet across all markets for this match
+            if best is not None:
+                edge,m,k,o,won = best
+                pnl = (o-1.0) if won else -1.0
+                sel["n"]+=1; sel["pnl"]+=pnl; sel["win"]+=won
+                sel["fold"][fi] = sel["fold"].get(fi,0.0)+pnl
+                d=sel_by_mkt[m]; d["n"]+=1; d["pnl"]+=pnl; d["win"]+=won
+        print(f"  fold {fi}: tested {te1-te0:,}")
+
+    def line(name,d):
+        n=d["n"]; roi=100*d["pnl"]/n if n else float('nan'); hit=100*d["win"]/n if n else float('nan')
+        return f"  {name:<10} bets={n:>6} hit={hit:>5.1f}% ROI={roi:>7.2f}% net={d['pnl']:>7.1f}u"
+
+    print("\n"+"="*70); print(f"(A) PER-MARKET value ROI  (best value pick in band [{args.lo},{args.hi}], margin {args.margin})"); print("="*70)
+    for m in sorted(per_market, key=lambda x:-(100*per_market[x]['pnl']/per_market[x]['n'] if per_market[x]['n'] else -99)):
+        print(line(m, per_market[m]))
+
+    print("\n"+"="*70); print("(B) CROSS-MARKET SELECTOR  (best value bet per match, all markets)"); print("="*70)
+    print(line("SELECTOR", sel))
+    folds_pos = sum(1 for v in sel["fold"].values() if v>0)
+    print(f"  folds positive: {folds_pos}/{len(sel['fold'])}")
+    print("  selector picks distributed across markets:")
+    for m in sorted(sel_by_mkt, key=lambda x:-sel_by_mkt[x]['n']):
+        if sel_by_mkt[m]["n"]>0: print("   "+line(m, sel_by_mkt[m]).strip())
+    print("\nREAD: a market/selector is a LEAD only if ROI>0, folds consistent, n large.")
+    print("Forward-validate with CLV before staking. Static CSV odds may overstate edge.")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/optimize_filters.py

@@ -0,0 +1,254 @@
+"""
+Filter Optimizer
+================
+Grid-search over filter thresholds (per market) using the existing
+diagnostic_backtest CSV. Finds the (confidence, edge, odds, reliability)
+combination that maximizes ROI while keeping bet volume reasonable.
+
+No re-prediction needed — pure offline simulation on the bets already
+captured. Output: per-market optimal thresholds + projected ROI lift +
+JSON patch ready to drop into config/market_thresholds.json.
+
+Usage:
+  python scripts/optimize_filters.py
+  python scripts/optimize_filters.py --csv reports/diagnostic_backtest_X.csv
+  python scripts/optimize_filters.py --min-bets 20 --apply
+"""
+
+import argparse
+import json
+import os
+import sys
+import glob
+import itertools
+from typing import List, Dict, Tuple, Optional
+import pandas as pd
+import numpy as np
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+AI_ENGINE_DIR = os.path.dirname(SCRIPT_DIR)
+sys.path.insert(0, AI_ENGINE_DIR)
+
+REPORTS_DIR = os.path.join(AI_ENGINE_DIR, "reports")
+CONFIG_PATH = os.path.join(AI_ENGINE_DIR, "config", "market_thresholds.json")
+
+
+def latest_csv() -> Optional[str]:
+    files = sorted(glob.glob(os.path.join(REPORTS_DIR, "diagnostic_backtest_*.csv")),
+                   key=os.path.getmtime, reverse=True)
+    return files[0] if files else None
+
+
+def load_backtest(path: str) -> pd.DataFrame:
+    df = pd.read_csv(path)
+    # Keep only playable + settled bets — these are what the SYSTEM
+    # actually placed and got an outcome on.
+    pdf = df[(df["playable"] == True) & (df["won"].notna())].copy()
+    pdf["won"] = pdf["won"].astype(bool)
+    pdf["calibrated_confidence"] = pdf["calibrated_confidence"].fillna(0)
+    pdf["ev_edge"] = pdf["ev_edge"].fillna(0)
+    pdf["odds"] = pdf["odds"].fillna(0)
+    pdf["odds_reliability"] = pdf["odds_reliability"].fillna(0.35)
+    return pdf
+
+
+def evaluate(pdf: pd.DataFrame, mask) -> Dict:
+    kept = pdf[mask]
+    if len(kept) == 0:
+        return {"n": 0, "hit_pct": 0, "profit": 0, "staked": 0, "roi_pct": 0}
+    wins = kept["won"].sum()
+    profit = kept["unit_profit"].sum()
+    staked = kept["stake_units"].sum()
+    return {
+        "n": int(len(kept)),
+        "hit_pct": round(100.0 * wins / len(kept), 2),
+        "profit": round(profit, 3),
+        "staked": round(staked, 3),
+        "roi_pct": round(100.0 * profit / staked, 2) if staked else 0,
+    }
+
+
+def grid_search_market(
+    market_df: pd.DataFrame,
+    market: str,
+    min_bets: int = 15,
+) -> List[Dict]:
+    """Try a wide grid of (min_conf, min_edge, max_edge, min_odds, max_odds,
+    min_reliability) combinations. Return all candidates with n >= min_bets,
+    sorted by ROI descending."""
+
+    conf_options = [0, 45, 50, 55, 60, 65, 70]
+    min_edge_options = [-1.0, -0.05, 0.0, 0.03, 0.05, 0.08]
+    max_edge_options = [10.0, 0.30, 0.20, 0.15, 0.10]
+    min_odds_options = [1.20, 1.30, 1.40, 1.50, 1.60, 1.80]
+    max_odds_options = [10.0, 3.0, 2.5, 2.2, 2.0]
+    rel_options = [0.0, 0.30, 0.45, 0.55]
+    consensus_options = ["any", "agree_or_null"]
+
+    candidates: List[Dict] = []
+    for mc, mine, maxe, mino, maxo, mrel, cons in itertools.product(
+        conf_options, min_edge_options, max_edge_options,
+        min_odds_options, max_odds_options, rel_options, consensus_options,
+    ):
+        if mine >= maxe or mino >= maxo:
+            continue
+        mask = (
+            (market_df["calibrated_confidence"] >= mc)
+            & (market_df["ev_edge"] >= mine)
+            & (market_df["ev_edge"] <= maxe)
+            & (market_df["odds"] >= mino)
+            & (market_df["odds"] <= maxo)
+            & (market_df["odds_reliability"] >= mrel)
+        )
+        if cons == "agree_or_null":
+            mask &= market_df["v27_consensus"] != "DISAGREE"
+        result = evaluate(market_df, mask)
+        if result["n"] >= min_bets:
+            candidates.append({
+                "market": market,
+                "min_conf": mc,
+                "min_edge": mine,
+                "max_edge": maxe,
+                "min_odds": mino,
+                "max_odds": maxo,
+                "min_reliability": mrel,
+                "consensus": cons,
+                **result,
+            })
+    candidates.sort(key=lambda r: (r["roi_pct"], r["n"]), reverse=True)
+    return candidates
+
+
+def baseline(pdf: pd.DataFrame, market: str) -> Dict:
+    m = pdf[pdf["market"] == market]
+    return evaluate(m, pd.Series([True] * len(m), index=m.index))
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--csv", default=None, help="Override CSV path")
+    parser.add_argument("--min-bets", type=int, default=15,
+                        help="Min bet count to consider a config valid")
+    parser.add_argument("--top-k", type=int, default=3,
+                        help="Show top K configs per market")
+    parser.add_argument("--apply", action="store_true",
+                        help="Patch config/market_thresholds.json with winners")
+    args = parser.parse_args()
+
+    csv_path = args.csv or latest_csv()
+    if not csv_path or not os.path.exists(csv_path):
+        print("No backtest CSV found.")
+        return
+    print(f"Loading: {csv_path}")
+    pdf = load_backtest(csv_path)
+    print(f"Playable + settled bets: {len(pdf)}")
+
+    markets = sorted(pdf["market"].dropna().unique())
+    print(f"Markets: {markets}\n")
+
+    all_winners: Dict[str, Dict] = {}
+
+    for market in markets:
+        market_df = pdf[pdf["market"] == market]
+        n_total = len(market_df)
+        base = baseline(pdf, market)
+
+        print(f"\n{'=' * 78}")
+        print(f"MARKET: {market}   (n={n_total} baseline_roi={base['roi_pct']}%)")
+        print(f"{'=' * 78}")
+
+        if n_total < args.min_bets * 2:
+            print(f"  Sample too small to grid-search reliably (n={n_total}). Skip.")
+            continue
+
+        candidates = grid_search_market(market_df, market, args.min_bets)
+        if not candidates:
+            print(f"  No config kept >= {args.min_bets} bets. Skip.")
+            continue
+
+        # Pareto-ish: show top-K by ROI but also one that keeps higher bet count
+        winners = candidates[:args.top_k]
+        keep_high_volume = None
+        for c in candidates:
+            if c["n"] >= max(40, n_total // 3) and c["roi_pct"] > base["roi_pct"]:
+                keep_high_volume = c
+                break
+
+        print(f"  {'rank':<5}{'n':>5}{'hit%':>7}{'roi%':>8}  "
+              f"{'min_conf':>9}{'min_edge':>10}{'max_edge':>10}"
+              f"{'min_odds':>10}{'max_odds':>10}{'min_rel':>9}{'cons':>15}")
+        for i, w in enumerate(winners, 1):
+            print(f"  {i:<5}{w['n']:>5}{w['hit_pct']:>7}{w['roi_pct']:>+8}"
+                  f"  {w['min_conf']:>9}{w['min_edge']:>+10.3f}{w['max_edge']:>+10.3f}"
+                  f"{w['min_odds']:>10.2f}{w['max_odds']:>10.2f}"
+                  f"{w['min_reliability']:>9.2f}{w['consensus']:>15}")
+        if keep_high_volume and keep_high_volume not in winners:
+            print(f"  high  {keep_high_volume['n']:>5}{keep_high_volume['hit_pct']:>7}"
+                  f"{keep_high_volume['roi_pct']:>+8}"
+                  f"  {keep_high_volume['min_conf']:>9}"
+                  f"{keep_high_volume['min_edge']:>+10.3f}"
+                  f"{keep_high_volume['max_edge']:>+10.3f}"
+                  f"{keep_high_volume['min_odds']:>10.2f}"
+                  f"{keep_high_volume['max_odds']:>10.2f}"
+                  f"{keep_high_volume['min_reliability']:>9.2f}"
+                  f"{keep_high_volume['consensus']:>15}")
+
+        # Pick a "good" recommendation: best ROI with n >= min_bets
+        # If best ROI is still negative, flag the market as unprofitable.
+        best = winners[0]
+        all_winners[market] = best
+        if best["roi_pct"] <= 0:
+            print(f"  ⚠️  Best config still loses money (ROI={best['roi_pct']}%) "
+                  f"— consider muting this market entirely.")
+        else:
+            print(f"  ✅  Best config: ROI={best['roi_pct']}% on {best['n']} bets "
+                  f"(vs baseline {base['roi_pct']}% on {n_total}).")
+
+    # ─── Aggregate impact ────────────────────────────────────────────────
+    print(f"\n{'=' * 78}")
+    print("AGGREGATE IMPACT (if we apply each market's best config)")
+    print(f"{'=' * 78}")
+    total_old_bets = total_old_profit = total_old_staked = 0
+    total_new_bets = total_new_profit = total_new_staked = 0
+    for market, win in all_winners.items():
+        base = baseline(pdf, market)
+        total_old_bets += base["n"]
+        total_old_profit += base["profit"]
+        total_old_staked += base["staked"]
+        total_new_bets += win["n"]
+        total_new_profit += win["profit"]
+        total_new_staked += win["staked"]
+    base_roi = 100.0 * total_old_profit / total_old_staked if total_old_staked else 0
+    new_roi = 100.0 * total_new_profit / total_new_staked if total_new_staked else 0
+    print(f"  Baseline:   {total_old_bets:>4} bets, "
+          f"profit={total_old_profit:+.2f}u, ROI={base_roi:+.2f}%")
+    print(f"  Optimized:  {total_new_bets:>4} bets, "
+          f"profit={total_new_profit:+.2f}u, ROI={new_roi:+.2f}%")
+    print(f"  Δ:          {total_new_bets - total_old_bets:+d} bets, "
+          f"{total_new_profit - total_old_profit:+.2f}u, "
+          f"{new_roi - base_roi:+.2f}pp")
+
+    # ─── Write JSON patch ────────────────────────────────────────────────
+    patch_path = os.path.join(REPORTS_DIR, "filter_optimization_patch.json")
+    patch = {market: {
+        "min_calibrated_confidence": win["min_conf"],
+        "min_ev_edge": win["min_edge"],
+        "max_ev_edge": win["max_edge"],
+        "min_odds": win["min_odds"],
+        "max_odds": win["max_odds"],
+        "min_odds_reliability": win["min_reliability"],
+        "require_v27_agree": win["consensus"] == "agree_or_null",
+        "expected_n_bets": win["n"],
+        "expected_hit_pct": win["hit_pct"],
+        "expected_roi_pct": win["roi_pct"],
+    } for market, win in all_winners.items()}
+    with open(patch_path, "w", encoding="utf-8") as f:
+        json.dump(patch, f, indent=2)
+    print(f"\nPatch saved: {patch_path}")
+
+    if args.apply:
+        print("\n--apply flag set. Patching not implemented yet — "
+              "review the patch JSON and update config/market_thresholds.json manually.")
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/test_score_coherence.py

@@ -0,0 +1,54 @@
+"""Smoke test for the score-coherence filter using the LAFC vs Sounders
+1-0 scenario from production. Verifies that markets that contradict the
+predicted score are correctly excluded from the coherent set, and that
+the markets the model got right are all included.
+"""
+import os, sys
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+from services.betting_brain import BettingBrain
+
+brain = BettingBrain()
+pkg = {
+    "score_prediction": {"ft": "1-0", "ht": "0-0"},
+}
+coh = brain._score_consistent_markets(pkg)
+print(f"Predicted: 1-0 (HT 0-0)")
+print(f"Coherent set size: {len(coh)}")
+print()
+
+# Each pick the system actually offered for the LAFC match, with whether
+# it was the *actual* winning pick.
+test_picks = [
+    ("MS", "1",        True,  "correct"),
+    ("MS", "2",        False, "wrong"),
+    ("MS", "X",        False, "wrong"),
+    ("DC", "1X",       True,  "correct"),
+    ("DC", "12",       True,  "correct"),
+    ("DC", "X2",       False, "wrong"),
+    ("OU25", "Üst",    False, "WRONG — system featured this"),
+    ("OU25", "Alt",    True,  "correct"),
+    ("OU35", "Alt",    True,  "correct"),
+    ("OU35", "Üst",    False, "wrong"),
+    ("BTTS", "Var",    False, "wrong"),
+    ("BTTS", "Yok",    True,  "correct"),
+    ("HT", "X",        True,  "correct"),
+    ("HT", "1",        False, "wrong"),
+    ("HTFT", "X/1",    True,  "correct"),
+    ("HTFT", "1/1",    False, "wrong (HT was 0-0)"),
+    ("HT_OU05", "Üst", False, "wrong"),
+    ("HT_OU05", "Alt", True,  "correct"),
+    ("OE", "Çift",     False, "wrong (1 is odd)"),
+    ("OE", "Tek",      True,  "correct"),
+]
+
+print(f"{'market':<10}{'pick':<10}{'real-win?':<12}{'in-coherent?':<14}{'match?'}")
+print("-" * 60)
+ok = 0
+for market, pick, would_win, note in test_picks:
+    in_coh = (market, pick) in coh
+    match = "✓" if in_coh == would_win else "✗ MISMATCH"
+    if in_coh == would_win: ok += 1
+    print(f"{market:<10}{pick:<10}{str(would_win):<12}{str(in_coh):<14}{match}  {note}")
+
+print()
+print(f"Result: {ok}/{len(test_picks)} picks correctly classified")

ai-engine/scripts/train_favorite_model.py

@@ -0,0 +1,112 @@
+"""
+Train Favorite-Policy Model (v1) — leak-free MS model for the validated strategy.
+================================================================================
+Trains a LEAK-FREE 1X2 model (drops the result-encoding columns) and saves it
+plus the feature list and policy metadata. This is the brain of the new system;
+the favourite-band value policy (odds ~1.5-2.2, model_prob>implied, flat stake)
+is applied on top of its probabilities at serving time.
+
+Honest holdout: trains on the first --holdout-frac of history, evaluates the
+EXACT policy on the most recent slice (never seen in training), then retrains
+on ALL history for the saved production artifact.
+
+Saves to models/favorite_v1/: model.json, feature_cols.json, metadata.json
+
+Usage: python scripts/train_favorite_model.py
+"""
+from __future__ import annotations
+import argparse, json, os, sys, datetime
+import numpy as np, pandas as pd, xgboost as xgb
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+OUT = os.path.join(AI_DIR, "models", "favorite_v1")
+
+META = {"match_id","home_team_id","away_team_id","league_id","mst_utc",
+        "score_home","score_away","ht_score_home","ht_score_away"}
+# Result-encoding leakage — never feed these to the model (train OR serve).
+LEAKY = {"home_goals_form","away_goals_form","total_goals","ht_total_goals",
+         "squad_diff","home_squad_quality","away_squad_quality",
+         "referee_home_bias","referee_avg_goals"}
+
+PARAMS = {"objective":"multi:softprob","num_class":3,"max_depth":5,"eta":0.05,
+          "subsample":0.8,"colsample_bytree":0.8,"tree_method":"hist","verbosity":0}
+
+
+def policy_eval(P, y, O, lo, hi, margin):
+    implied = np.where(O > 1.0, 1.0/O, np.nan)
+    edge = np.where(np.isnan(implied), -9.0, P - implied)
+    pick = edge.argmax(1); pe = edge[np.arange(len(y)), pick]; po = O[np.arange(len(y)), pick]
+    bet = (pe > margin) & (po >= lo) & (po < hi)
+    win = (pick == y) & bet
+    pnl = np.where(win, po-1.0, -1.0)[bet]
+    n = int(bet.sum())
+    return {"bets": n, "hit_pct": round(100*win.sum()/max(n,1),1),
+            "roi_pct": round(100*pnl.sum()/max(n,1),2), "net_u": round(float(pnl.sum()),1)}
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--lo", type=float, default=1.5)
+    ap.add_argument("--hi", type=float, default=2.2)
+    ap.add_argument("--margin", type=float, default=0.0)
+    ap.add_argument("--holdout-frac", type=float, default=0.15)
+    ap.add_argument("--estimators", type=int, default=300)
+    args = ap.parse_args()
+
+    print(f"Loading {CSV} ...")
+    df = pd.read_csv(CSV, low_memory=False).sort_values("mst_utc").reset_index(drop=True)
+    sh = pd.to_numeric(df["score_home"], errors="coerce")
+    sa = pd.to_numeric(df["score_away"], errors="coerce")
+    ok = sh.notna() & sa.notna()
+    df, sh, sa = df[ok].reset_index(drop=True), sh[ok.values].values, sa[ok.values].values
+    y = np.where(sh > sa, 0, np.where(sh == sa, 1, 2))
+    O = df[["odds_ms_h","odds_ms_d","odds_ms_a"]].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    feats = [c for c in df.columns if c not in META and not c.startswith("label_") and c not in LEAKY]
+    X = df[feats].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    print(f"  {len(df):,} rows, {len(feats)} leak-free features")
+
+    # ── Honest holdout (last slice, never trained on) ──
+    cut = int(len(df) * (1 - args.holdout_frac))
+    bst = xgb.train(PARAMS, xgb.DMatrix(X[:cut], label=y[:cut]), num_boost_round=args.estimators)
+    Ph = bst.predict(xgb.DMatrix(X[cut:]))
+    acc = float((Ph.argmax(1) == y[cut:]).mean())
+    hold = policy_eval(Ph, y[cut:], O[cut:], args.lo, args.hi, args.margin)
+    print(f"\nHOLDOUT (last {args.holdout_frac:.0%}, {len(df)-cut:,} matches, never seen):")
+    print(f"  MS accuracy: {acc*100:.1f}%")
+    print(f"  POLICY band[{args.lo},{args.hi}] margin {args.margin}: {hold}")
+
+    # ── Production model: retrain on ALL history ──
+    print("\nTraining production model on ALL history ...")
+    final = xgb.train(PARAMS, xgb.DMatrix(X, label=y), num_boost_round=args.estimators)
+    os.makedirs(OUT, exist_ok=True)
+    final.save_model(os.path.join(OUT, "model.json"))
+    with open(os.path.join(OUT, "feature_cols.json"), "w", encoding="utf-8") as f:
+        json.dump(feats, f, ensure_ascii=False, indent=2)
+    meta = {
+        "version": "favorite_v1",
+        "trained_at": datetime.datetime.now().isoformat(timespec="seconds"),
+        "market": "MS",
+        "classes": {"0": "home(1)", "1": "draw(X)", "2": "away(2)"},
+        "policy": {"odds_lo": args.lo, "odds_hi": args.hi, "margin": args.margin,
+                   "stake": "flat 1u", "rule": "bet model's max value edge if picked odds in band",
+                   "never": ["longshots odds>=hi", "parlays/combos"]},
+        "n_train": len(df), "n_features": len(feats),
+        "leaky_excluded": sorted(LEAKY),
+        "holdout_eval": {"accuracy_pct": round(acc*100,1), **hold},
+        "caveat": "CSV odds are a static capture, not verified closing. Forward paper-trade with real CLV before staking.",
+    }
+    with open(os.path.join(OUT, "metadata.json"), "w", encoding="utf-8") as f:
+        json.dump(meta, f, ensure_ascii=False, indent=2)
+    print(f"\n✅ Saved production model to {OUT}/")
+    print(f"   model.json, feature_cols.json ({len(feats)} feats), metadata.json")
+    print("\nNEXT: serving wrapper that loads this + applies the policy to upcoming")
+    print("matches, logs paper-trade picks, and we measure real forward CLV/ROI.")
+
+
+if __name__ == "__main__":
+    main()

ai-engine/scripts/walkforward_oddsblind.py

@@ -0,0 +1,191 @@
+"""
+Walk-Forward Odds-Blind Experiment — THE pivotal test.
+======================================================
+Question this answers: can a model BEAT THE MARKET out-of-sample, betting only
+on information the price doesn't already contain?
+
+Method (no leakage, time-ordered):
+  * data sorted by kickoff (mst_utc); train on the past, test on the future,
+    rolled over several folds.
+  * TWO models on the MS (1X2) market:
+       ALL    = every feature INCLUDING the bookmaker odds (what the live
+                engine does -> it mostly re-learns the price).
+       BLIND  = identical but odds/implied/_present columns REMOVED, so the
+                model must disagree with the market using fundamentals only.
+  * For each, an honest value-bet simulation on the test fold using the REAL
+    odds payouts (margin included): bet the outcome with the biggest
+    model_prob - implied_prob edge above a margin; ROI = realized P/L per 1u.
+
+Read: if BLIND's value ROI is consistently > 0 across folds, there is a real,
+exploitable lead. If both are <= 0 (expected), these markets aren't beatable
+with this data and the honest move is to stop staking.
+
+Usage:
+  python scripts/walkforward_oddsblind.py
+  python scripts/walkforward_oddsblind.py --folds 6 --estimators 300
+"""
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+import numpy as np
+import pandas as pd
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try:
+        sys.stdout.reconfigure(encoding="utf-8")
+    except Exception:
+        pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+
+import xgboost as xgb  # noqa: E402
+
+META = {"match_id", "home_team_id", "away_team_id", "league_id", "mst_utc",
+        "score_home", "score_away", "ht_score_home", "ht_score_away"}
+
+# Confirmed target leakage: *_goals_form integer-valued and ~0.63 correlated
+# with THIS match's goals; their diff equals the actual goal diff 73% of the
+# time. Excluded so the experiment measures genuine pre-match predictive power.
+LEAKY = {
+    # CONFIRMED (encode the actual match result):
+    "home_goals_form", "away_goals_form",  # ~0.63 corr w/ this match's goals
+    "total_goals",                          # this match's full-time total
+    "ht_total_goals",                       # this match's half-time total
+    # STRONG SUSPECTS (dominate importance + high outcome corr; audit extractor):
+    "squad_diff", "home_squad_quality", "away_squad_quality",
+    "referee_home_bias", "referee_avg_goals",
+}
+
+
+def is_odds_col(c: str) -> bool:
+    cl = c.lower()
+    return ("odds" in cl) or ("implied" in cl)
+
+
+def logloss(y: np.ndarray, p: np.ndarray) -> float:
+    p = np.clip(p, 1e-9, 1 - 1e-9)
+    return float(-np.mean(np.log(p[np.arange(len(y)), y])))
+
+
+def value_sim(proba: np.ndarray, y: np.ndarray, odds: np.ndarray,
+              margin: float) -> dict:
+    """Bet the class with the biggest (model_prob - 1/odds) edge above margin."""
+    implied = np.where(odds > 1.0, 1.0 / odds, np.nan)
+    edge = proba - implied
+    # ignore classes without valid odds
+    edge = np.where(np.isnan(implied), -9.0, edge)
+    pick = np.argmax(edge, axis=1)
+    best_edge = edge[np.arange(len(y)), pick]
+    bet = best_edge > margin
+    n = int(bet.sum())
+    if n == 0:
+        return {"n": 0, "roi": None, "hit": None}
+    win = (pick == y) & bet
+    pick_odds = odds[np.arange(len(y)), pick]
+    pnl = np.where(win, pick_odds - 1.0, -1.0)
+    pnl = pnl[bet]
+    return {"n": n, "roi": round(100.0 * pnl.sum() / n, 2),
+            "hit": round(100.0 * win[bet].sum() / n, 1)}
+
+
+def train_eval(Xtr, ytr, Xte, yte, odds_te, est, margins):
+    dtr = xgb.DMatrix(Xtr, label=ytr)
+    dte = xgb.DMatrix(Xte)
+    params = {"objective": "multi:softprob", "num_class": 3, "max_depth": 5,
+              "eta": 0.05, "subsample": 0.8, "colsample_bytree": 0.8,
+              "tree_method": "hist", "verbosity": 0}
+    booster = xgb.train(params, dtr, num_boost_round=est)
+    proba = booster.predict(dte)
+    out = {"logloss": round(logloss(yte, proba), 4),
+           "acc": round(100.0 * (proba.argmax(1) == yte).mean(), 1)}
+    for mg in margins:
+        out[f"val@{mg}"] = value_sim(proba, yte, odds_te, mg)
+    return out
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--folds", type=int, default=5)
+    ap.add_argument("--estimators", type=int, default=250)
+    ap.add_argument("--test-frac", type=float, default=0.5,
+                    help="Fraction at the end used as rolling OOS (default 0.5)")
+    args = ap.parse_args()
+
+    print(f"Loading {CSV} ...")
+    df = pd.read_csv(CSV, low_memory=False)
+    df = df.sort_values("mst_utc").reset_index(drop=True)
+    print(f"  {len(df)} rows, {df.shape[1]} cols")
+
+    # Derive true MS outcome from scores: 0=home,1=draw,2=away (robust, no label trust)
+    sh = pd.to_numeric(df["score_home"], errors="coerce")
+    sa = pd.to_numeric(df["score_away"], errors="coerce")
+    y = np.where(sh > sa, 0, np.where(sh == sa, 1, 2))
+    valid = sh.notna() & sa.notna()
+    df, y = df[valid].reset_index(drop=True), y[valid.values]
+
+    odds = df[["odds_ms_h", "odds_ms_d", "odds_ms_a"]].apply(
+        pd.to_numeric, errors="coerce").fillna(0.0).values
+
+    feat_all = [c for c in df.columns if c not in META and not c.startswith("label_")
+                and c not in LEAKY]
+    feat_blind = [c for c in feat_all if not is_odds_col(c)]
+    print(f"  excluded leaky cols: {sorted(LEAKY)}")
+    Xall = df[feat_all].apply(pd.to_numeric, errors="coerce").fillna(0.0)
+    Xblind = df[feat_blind].apply(pd.to_numeric, errors="coerce").fillna(0.0)
+    print(f"  features: ALL={len(feat_all)}  BLIND={len(feat_blind)} "
+          f"(dropped {len(feat_all)-len(feat_blind)} odds cols)")
+    print(f"  base rates: home={100*(y==0).mean():.1f}% draw={100*(y==1).mean():.1f}% "
+          f"away={100*(y==2).mean():.1f}%")
+
+    n = len(df)
+    start = int(n * (1 - args.test_frac))
+    bounds = np.linspace(start, n, args.folds + 1, dtype=int)
+    margins = [0.0, 0.05, 0.10]
+
+    agg = {"ALL": {f"val@{m}": [] for m in margins}, "BLIND": {f"val@{m}": [] for m in margins}}
+    agg["ALL"]["logloss"] = []; agg["BLIND"]["logloss"] = []
+
+    print(f"\nWalk-forward: {args.folds} folds, train=expanding, est={args.estimators}\n")
+    hdr = f"{'fold':<5}{'model':<7}{'logloss':>9}{'acc%':>7}" + "".join(
+        f"{('val@'+str(m)):>22}" for m in margins)
+    print(hdr); print("-" * len(hdr))
+    for i in range(args.folds):
+        te0, te1 = bounds[i], bounds[i + 1]
+        if te1 - te0 < 50:
+            continue
+        tr = slice(0, te0)
+        te = slice(te0, te1)
+        for name, X in (("ALL", Xall), ("BLIND", Xblind)):
+            r = train_eval(X.iloc[tr].values, y[tr], X.iloc[te].values, y[te],
+                           odds[te], args.estimators, margins)
+            agg[name]["logloss"].append(r["logloss"])
+            cells = ""
+            for m in margins:
+                v = r[f"val@{m}"]
+                agg[name][f"val@{m}"].append(v)
+                cells += f"{('n=' + str(v['n']) + ' roi=' + str(v['roi'])):>22}"
+            print(f"{i:<5}{name:<7}{r['logloss']:>9}{r['acc']:>7}{cells}")
+        print()
+
+    print("=" * 70)
+    print("AGGREGATE (sum bets, weighted ROI across folds)")
+    print("=" * 70)
+    for name in ("ALL", "BLIND"):
+        ll = np.mean(agg[name]["logloss"]) if agg[name]["logloss"] else float("nan")
+        print(f"\n{name}  mean logloss={ll:.4f}")
+        for m in margins:
+            vs = agg[name][f"val@{m}"]
+            tot_n = sum(v["n"] for v in vs)
+            tot_pnl = sum((v["roi"] / 100.0 * v["n"]) for v in vs if v["roi"] is not None)
+            roi = round(100.0 * tot_pnl / tot_n, 2) if tot_n else None
+            print(f"   margin {m}: total_bets={tot_n:>6}  ROI(flat1u)={roi}%")
+    print("\nREAD: BLIND ROI>0 across margins/folds = real edge. Both <=0 = no")
+    print("exploitable edge in MS with this data (stop staking; the -EV is the vig).")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

ai-engine/services/betting_brain.py

@@ -12,16 +12,52 @@ from typing import Any, Dict, List, Optional, Tuple
 
 class BettingBrain:
     MIN_ODDS = 1.30
-    MIN_BET_SCORE = 72.0
-    MIN_WATCH_SCORE = 62.0
+    MIN_BET_SCORE = 62.0
+    MIN_WATCH_SCORE = 52.0
     MIN_BAND_SAMPLE = 8
     HARD_DIVERGENCE = 0.22
     SOFT_DIVERGENCE = 0.14
     EXTREME_MODEL_PROB = 0.85
     EXTREME_GAP = 0.30
     SNIPER_BYPASSABLE_VETOES = {"play_score_too_low"}
+    # V31d: value-tier underdogs are bet on the odds-premium edge, NOT on
+    # high win-probability. These two vetoes encode a favorite-picking rule
+    # (demand >45% confidence) that structurally excludes every profitable
+    # underdog, so we waive them when a row matches an MS value tier.
+    # Genuine safety vetoes (extreme_neg_ev, ev_too_high_trap, htft_reversal
+    # _risk_high, v25_v27_hard_disagreement, low_reliability_hard_block) are
+    # NOT in this set and still reject.
+    VALUE_TIER_BYPASSABLE_VETOES = {"calibrated_confidence_too_low", "play_score_too_low"}
+    VALUE_TIER_NAMES = {"premium", "strong", "standard"}
+    # V31d: value-regime score floors (replaces favorite-confidence scoring
+    # for value-tier matches). premium clears MIN_BET_SCORE(62) → BET;
+    # strong/standard are capped below it → WATCH (visible, not staked)
+    # because the 60-day data shows those bands break even.
+    VALUE_TIER_BASE_SCORE = {"premium": 70.0, "strong": 56.0, "standard": 54.0}
+    # V31d: flat, small stake for value-tier underdogs. Hit rate ~20% with
+    # long losing streaks (60d: up to 35 in a row) — the edge is in FREQUENCY,
+    # not per-bet size. Keep stake small to survive variance. Tunable: raise
+    # only if the bettor's bankroll/risk appetite allows deeper drawdowns.
+    VALUE_TIER_STAKE_UNITS = 0.5
     TRAP_MARKET_GAP = 0.10
 
+    # ── V31f: NATIONAL-TEAM REGIME ───────────────────────────────────────
+    # National matches behave nothing like clubs (2300-match backtest):
+    #   * Only MS carries edge — OU/BTTS/HT/DC/OE all -12%..-21% → hard mute.
+    #   * MS edge lives in the 4.0–7.0 odds band for HAZIRLIK/ELEME fixtures
+    #     (+17% ROI, stable across older/newer halves: +22%/+24%).
+    #   * Favorites (odds<3) lose (-10..-18%); TURNUVA inverts the pattern
+    #     (4-7 band is -9% there) → tournaments get NO bet (analysis only).
+    # Calibration is fine; this is a *bet-selection* gate, applied only when
+    # match_info.is_national is True. Clubs are completely unaffected.
+    # See mds/national-team-strategy.md.
+    NATIONAL_BET_MARKET = "MS"
+    NATIONAL_MIN_ODDS = 4.0
+    NATIONAL_MAX_ODDS = 7.0
+    NATIONAL_ALLOWED_COMPETITIONS = {"HAZIRLIK", "ELEME"}
+    NATIONAL_BASE_SCORE = 66.0          # clears MIN_BET_SCORE(62) when gate passes
+    NATIONAL_STAKE_UNITS = 0.5          # flat, high-variance band (~24% hit)
+
     MARKET_MIN_CONFIDENCE = {
         "MS": 45.0,
         "DC": 55.0,
@@ -39,6 +75,181 @@ class BettingBrain:
 
     SNIPER_BLOCKED_MARKETS = {"HT", "HTFT", "OE", "CARDS", "HT_OU05", "HT_OU15"}
 
+    # V30: NO markets muted — backtest tüm marketlerin gerçek ROI'sini görmeli.
+    # Tier sistemi zaten filtreleme yapıyor; mute etmek veri kaybına yol açar.
+    MUTED_MARKETS = set()
+
+    # ═══════════════════════════════════════════════════════════════════
+    # V31d: KANITA DAYALI KADEMELİ DEĞER SİSTEMİ (Evidence-Based Tiers)
+    # ═══════════════════════════════════════════════════════════════════
+    # User directive: show 3 quality levels so the bettor picks by risk
+    # appetite ("hangi bahis hangi oranda tutuyor"). Each MS underdog tier
+    # carries a `value_tier` label that propagates to the UI. Only the
+    # PREMIUM band is auto-staked (BET); strong/standard surface as WATCH
+    # (full analysis shown, not staked) because the data says they break
+    # even — see new_gate_sim.py.
+    #
+    # Validated on 60-day, 72,582-settled-row multi-pick backtest
+    # (ms_envelope.py + new_gate_sim.py, span 2026-04-17..05-28):
+    #   PREMIUM  (6.0-7.5, gap≥0, protective vetoes kept):
+    #       602 bets, +32.7% ROI, +39.4u, 20.6% hit, avgOdd 6.50
+    #       ALL 6 weeks positive (+13.7%..+52.9%); OOS(>05-24) +47.4%;
+    #       survives dropping top-5 wins (+24%). 14.3 bets/day.
+    #   STRONG   (5.0-6.0, gap≥0): ~breakeven (-1%) → WATCH, not staked.
+    #   STANDARD (3.0-5.0, gap≥0): +0.5% breakeven → WATCH, volume zone.
+    #
+    # WHY 6.0-7.5 (not 6.0-50.0 as in V31c): the edge is concentrated.
+    #   odds 6.0-7.0 +35% | 7.0-8.0 ~breakeven | 8.0+ NEGATIVE (longshot
+    #   graveyard, -10..-26% ROI). The old wide premium tier let losing
+    #   longshots in. Above 7.5 the model's edge evaporates.
+    #
+    # ROOT-CAUSE FIX (the volume crisis): underdogs were structurally
+    # un-bettable. Two vetoes (calibrated_confidence_too_low,
+    # play_score_too_low) auto-REJECTED every dog because they demand
+    # >45% model confidence — a FAVORITE-picking rule. A 6.5 dog wins
+    # ~20% of the time; that IS the edge (odds premium), not a defect.
+    # For value-tier matches we bypass those two vetoes and score from
+    # the validated tier quality instead of win-probability. Genuine
+    # protections stay: extreme_neg_ev, ev_too_high_trap, htft_reversal
+    # _risk_high, v25_v27_hard_disagreement. Result: 28 → 602 staked
+    # bets (22x volume), -1.6u → +39.4u profit. ALL rich analysis data
+    # (market_board, v25/v27, triple_value, probs) is untouched — only
+    # the `playable` flag changes.
+    #
+    # MULTI-LEG VERDICT (definitive): parlays DESTROY edge.
+    #   1-leg +3.4% → 2-leg -32% → 3-leg -67% → 4-leg -83%.
+    #   System must bet SINGLES only. No combo recommendations.
+    #
+    # Non-MS markets: ultrastrict tiers (rarely pass BET) → info-only.
+    # All non-MS configurations showed negative ROI in backtest.
+    # ═══════════════════════════════════════════════════════════════════
+    MARKET_ODDS_TIERS = {
+        # ── MS (Match Score / 1X2) — the ONLY profitable market ────────
+        "MS": [
+            # PREMIUM — the validated edge. 6.0-7.5 odds, model >= market.
+            # 60d: 602 bets, +32.7% ROI, all weeks positive. AUTO-STAKED.
+            # Low hit (~20%) → high variance; stake stays small (see _brain_stake).
+            {"min_odds": 6.00, "max_odds": 7.50, "min_edge": -0.20,
+             "max_edge": 0.25, "min_reliability": 0.30,
+             "min_model_gap": 0.0,
+             "require_v27_agree": False, "require_no_trap": False,
+             "value_tier": "premium",
+             "label": "ms_underdog_premium"},
+
+            # STRONG — 5.0-6.0. Breakeven (-1%) → WATCH (visible, not staked).
+            {"min_odds": 5.00, "max_odds": 6.00, "min_edge": -0.20,
+             "max_edge": 0.25, "min_reliability": 0.30,
+             "min_model_gap": 0.0,
+             "require_v27_agree": False, "require_no_trap": False,
+             "value_tier": "strong",
+             "label": "ms_underdog_strong"},
+
+            # STANDARD — 3.0-5.0 volume zone. +0.5% breakeven → WATCH.
+            {"min_odds": 3.00, "max_odds": 5.00, "min_edge": -0.18,
+             "max_edge": 0.25, "min_reliability": 0.30,
+             "min_model_gap": 0.0,
+             "require_v27_agree": False, "require_no_trap": False,
+             "value_tier": "standard",
+             "label": "ms_underdog_standard"},
+        ],
+
+        # ── Non-MS markets: visible but NOT playable ───────────────────
+        # All non-MS markets showed negative ROI in 50K-row backtest.
+        # Tiers exist so the model's read is surfaced (bet_summary),
+        # but criteria are strict enough that almost nothing passes BET.
+        # The user sees info; the system doesn't lose money on them.
+
+        "DC": [
+            {"min_odds": 1.15, "max_odds": 1.60, "min_edge": 0.02,
+             "max_edge": 0.12, "min_reliability": 0.55,
+             "max_model_gap": -0.02,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "dc_ultrastrict"},
+        ],
+
+        "OU25": [
+            {"min_odds": 1.60, "max_odds": 2.20, "min_edge": 0.02,
+             "max_edge": 0.10, "min_reliability": 0.55,
+             "max_model_gap": -0.03,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "ou25_ultrastrict"},
+        ],
+
+        "OU35": [
+            {"min_odds": 1.50, "max_odds": 2.50, "min_edge": 0.02,
+             "max_edge": 0.12, "min_reliability": 0.50,
+             "max_model_gap": -0.02,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "ou35_ultrastrict"},
+        ],
+
+        "BTTS": [
+            {"min_odds": 1.60, "max_odds": 2.10, "min_edge": 0.02,
+             "max_edge": 0.10, "min_reliability": 0.55,
+             "max_model_gap": -0.03,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "btts_ultrastrict"},
+        ],
+
+        "HT": [
+            {"min_odds": 2.00, "max_odds": 3.50, "min_edge": 0.02,
+             "max_edge": 0.12, "min_reliability": 0.50,
+             "max_model_gap": -0.02,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "ht_ultrastrict"},
+        ],
+
+        "OU15": [
+            {"min_odds": 1.30, "max_odds": 2.00, "min_edge": 0.02,
+             "max_edge": 0.12, "min_reliability": 0.50,
+             "max_model_gap": -0.02,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "ou15_ultrastrict"},
+        ],
+
+        "HTFT": [
+            {"min_odds": 4.00, "max_odds": 15.00, "min_edge": 0.03,
+             "max_edge": 0.15, "min_reliability": 0.45,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "htft_ultrastrict"},
+        ],
+
+        "OE": [
+            {"min_odds": 1.80, "max_odds": 2.10, "min_edge": 0.02,
+             "max_edge": 0.08, "min_reliability": 0.55,
+             "max_model_gap": -0.03,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "oe_ultrastrict"},
+        ],
+
+        "HT_OU05": [
+            {"min_odds": 1.30, "max_odds": 2.00, "min_edge": 0.02,
+             "max_edge": 0.12, "min_reliability": 0.50,
+             "max_model_gap": -0.02,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "ht_ou05_ultrastrict"},
+        ],
+
+        "HT_OU15": [
+            {"min_odds": 1.60, "max_odds": 3.00, "min_edge": 0.02,
+             "max_edge": 0.12, "min_reliability": 0.50,
+             "max_model_gap": -0.02,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "ht_ou15_ultrastrict"},
+        ],
+
+        "CARDS": [
+            {"min_odds": 1.60, "max_odds": 2.50, "min_edge": 0.02,
+             "max_edge": 0.10, "min_reliability": 0.50,
+             "max_model_gap": -0.02,
+             "require_v27_agree": False, "require_no_trap": True,
+             "label": "cards_ultrastrict"},
+        ],
+    }
+
+    # Legacy flat envelope (backward compat for markets not in tiered system)
+    MARKET_OPTIMAL_FILTERS = {}
+
     MARKET_PRIORS = {
         "DC": 4.0,
         "OU15": 3.0,
@@ -52,7 +263,14 @@ class BettingBrain:
         "OE": -12.0,
     }
 
-    def judge(self, package: Dict[str, Any]) -> Dict[str, Any]:
+    def judge(
+        self,
+        package: Dict[str, Any],
+        ms_real_odds: Optional[Dict[str, float]] = None,
+    ) -> Dict[str, Any]:
+        # V35c: real bookmaker MS odds (from odds_data) for reference rows —
+        # the brain must never display synthetic 1/p "fair odds" as offered.
+        self._ms_real_odds = ms_real_odds if isinstance(ms_real_odds, dict) else {}
         v27_engine = package.get("v27_engine")
         if not isinstance(v27_engine, dict):
             return package
@@ -86,6 +304,36 @@ class BettingBrain:
         watchlist.sort(key=self._candidate_sort_key, reverse=True)
         no_value.sort(key=self._candidate_sort_key, reverse=True)
 
+        # ── SCORE COHERENCE FILTER ──────────────────────────────────────
+        # If the model also produced a score prediction (e.g. 1-0), pick
+        # main_pick from the subset of candidates that would WIN at that
+        # score. Stops the system from recommending OU25 Üst while also
+        # predicting 1-0 (only 1 goal). Falls back to original list if no
+        # coherent candidate exists.
+        coherent_set = self._score_consistent_markets(guarded)
+        coherent_flag = False
+        if coherent_set:
+            def is_coherent(row: Dict[str, Any]) -> bool:
+                m = str(row.get("market") or "")
+                p = str(row.get("pick") or "")
+                return (m, p) in coherent_set
+
+            approved_coh = [r for r in approved if is_coherent(r)]
+            watchlist_coh = [r for r in watchlist if is_coherent(r)]
+
+            if approved_coh:
+                approved = approved_coh
+                coherent_flag = True
+            elif watchlist_coh:
+                # No coherent BET candidates — at least promote a coherent
+                # watch over an incoherent BET.
+                watchlist = watchlist_coh + [r for r in watchlist if not is_coherent(r)]
+                coherent_flag = True
+            # Tag every row so the UI/diagnostics can see what happened
+            for row in judged_rows.values():
+                row.setdefault("betting_brain", {})
+                row["betting_brain"]["score_coherent"] = is_coherent(row)
+
         original_main = guarded.get("main_pick") or {}
         main_pick = None
         decision = "NO_BET"
@@ -142,10 +390,11 @@ class BettingBrain:
 
         rejected = [d for d in decisions if d.get("action") == "REJECT"]
         guarded["betting_brain"] = {
-            "version": "judge-v1",
+            "version": "judge-v31f-national-regime",
             "decision": decision,
             "reason": decision_reason,
             "main_pick_key": main_key or None,
+            "score_coherent_filter_applied": coherent_flag,
             "approved_count": len(approved),
             "watchlist_count": len(watchlist),
             "rejected_count": len(rejected),
@@ -171,6 +420,10 @@ class BettingBrain:
         pick = str(row.get("pick") or "")
         model_prob = self._market_probability(row, package)
         odds = self._safe_float(row.get("odds"), 0.0) or 0.0
+        # V31f: national-team match flags (set by orchestrator in match_info).
+        _mi = package.get("match_info") or {}
+        is_national = bool(_mi.get("is_national"))
+        competition_type = str(_mi.get("competition_type") or "")
         implied = (1.0 / odds) if odds > 1.0 else 0.0
         model_gap = (model_prob - implied) if model_prob is not None and implied > 0 else None
         calibrated_conf = self._safe_float(row.get("calibrated_confidence", row.get("confidence")), 0.0) or 0.0
@@ -184,6 +437,21 @@ class BettingBrain:
         triple_is_value = bool((triple or {}).get("is_value"))
         consensus = str((package.get("v27_engine") or {}).get("consensus") or "").upper()
 
+        # V29c: Compute trap_market_flag early (needed by tier require_no_trap)
+        trap_market_flag = False
+        trap_market_gap = None
+        if isinstance(triple, dict):
+            _band_rate = self._safe_float(triple.get("band_rate"))
+            _implied = self._safe_float(triple.get("implied_prob"))
+            if (
+                _band_rate is not None
+                and _implied is not None
+                and band_sample >= self.MIN_BAND_SAMPLE
+                and (_implied - _band_rate) > self.TRAP_MARKET_GAP
+            ):
+                trap_market_flag = True
+                trap_market_gap = round(_implied - _band_rate, 4)
+
         positives: List[str] = []
         issues: List[str] = []
         vetoes: List[str] = []
@@ -200,7 +468,7 @@ class BettingBrain:
         if market in self.SNIPER_BLOCKED_MARKETS:
             is_value_sniper = False
         if is_value_sniper:
-            score += 20.0
+            score += 8.0  # V29b: reduced from 20, tiers do the real filtering
             positives.append("value_sniper_override")
 
         score += max(0.0, min(20.0, calibrated_conf * 0.22))
@@ -220,7 +488,7 @@ class BettingBrain:
         if odds_rel < 0.30:
             score -= 22.0
             issues.append("very_low_reliability_league")
-            if market in {"MS", "DC", "OU25", "BTTS"} and not is_value_sniper:
+            if market in {"MS", "DC", "OU25", "BTTS"}:  # V29: hard veto, no sniper bypass
                 vetoes.append("low_reliability_league_hard_block")
         elif odds_rel < 0.45:
             score -= 12.0
@@ -243,6 +511,85 @@ class BettingBrain:
         if play_score < 50.0 and not is_value_sniper:
             vetoes.append("play_score_too_low")
 
+        # ── HARD EV-EDGE VETO ───────────────────────────────────────────
+        # Diagnostic backtest (1000 maç, 524 settled bet) gösterdi ki
+        # ev_edge < 0 olan bahisler %76 of all picks ve ROI yaklaşık -%16.
+        # ev_edge < 0 = "model market'in altında olasılık veriyor" = vig'i
+        # yiyemeyeceğimiz negative-EV bahis. Hard veto: oynama.
+        # Sniper override hâlâ geçer (yüksek convicted alternatif pick'ler).
+        # V29b: negative_ev_edge hard veto REMOVED — tier system handles
+        # edge bounds per-market via min_edge. MS underdog tier allows
+        # ev >= -0.15, so a universal ev<0 veto would kill profitable bets.
+        if ev_edge < -0.20:  # Only veto truly extreme negative edge
+            vetoes.append("extreme_negative_ev_edge")
+            issues.append(f"ev_edge={ev_edge:.3f}_extreme_negative")
+        # Trap edge: bizim diagnostic backtest'te ev_edge >= 0.20 olan tüm
+        # bahisler kaybediyordu (n=10, hepsi -%25+ ROI). Model market'i bu
+        # kadar yanlış buluyorsa muhtemelen modelin kendisinin yanlış olduğu
+        # bir senaryo (eksik info, tuhaf maç, vs.) — oynama.
+        if ev_edge >= 0.30:  # V29b: raised from 0.20, tiers cap at 0.25
+            vetoes.append("ev_edge_too_high_trap")
+            issues.append(f"ev_edge={ev_edge:.3f}_trap_range")
+
+        # ── MUTED MARKETS (grid search showed no profitable config) ──
+        if market in self.MUTED_MARKETS:  # V29: hard veto, no sniper bypass
+            vetoes.append("market_muted_by_backtest")
+            issues.append(f"market_{market}_muted")
+
+        # ── V30: ODDS-TIERED ENVELOPE (from 7K backtest grid search) ──
+        # Each market has multiple odds zones with different filters.
+        # If a bet doesn't fit ANY tier, it gets vetoed.
+        # V30: added model_gap filtering — data shows model>market is
+        # inversely correlated with winning for BTTS/OU25.
+        tiers = self.MARKET_ODDS_TIERS.get(market, [])
+        # Also check legacy flat envelope for backward compat
+        legacy_env = self.MARKET_OPTIMAL_FILTERS.get(market)
+        tier_matched = False
+        tier_label = None
+        tier_value = None  # V31c: quality tier (premium/strong/standard)
+        if tiers:
+            for tier in tiers:
+                if not (tier["min_odds"] <= odds <= tier["max_odds"]):
+                    continue
+                if ev_edge < tier["min_edge"] or ev_edge > tier["max_edge"]:
+                    continue
+                if odds_rel < tier["min_reliability"]:
+                    continue
+                if tier.get("require_v27_agree") and consensus != "AGREE":
+                    continue
+                if tier.get("require_no_trap") and trap_market_flag:
+                    continue
+                # V30: model-market gap filter
+                if model_gap is not None:
+                    if "min_model_gap" in tier and model_gap < tier["min_model_gap"]:
+                        continue
+                    if "max_model_gap" in tier and model_gap > tier["max_model_gap"]:
+                        continue
+                tier_matched = True
+                tier_label = tier.get("label")
+                tier_value = tier.get("value_tier")  # V31c
+                break
+            if not tier_matched:
+                vetoes.append("outside_all_odds_tiers")
+                issues.append(f"no_profitable_tier_for_{market}_at_odds_{odds:.2f}")
+        elif legacy_env:
+            if ev_edge < legacy_env["min_edge"]:
+                vetoes.append("outside_envelope_edge_low")
+            if ev_edge > legacy_env["max_edge"]:
+                vetoes.append("outside_envelope_edge_high")
+            if odds and odds < legacy_env["min_odds"]:
+                vetoes.append("outside_envelope_odds_low")
+            if odds and odds > legacy_env["max_odds"]:
+                vetoes.append("outside_envelope_odds_high")
+            if odds_rel < legacy_env["min_reliability"]:
+                vetoes.append("outside_envelope_reliability_low")
+            if legacy_env.get("require_v27_agree") and consensus != "AGREE":
+                vetoes.append("outside_envelope_v27_must_agree")
+
+        # V31d: a matched value tier is the validated profitable signal.
+        # It unlocks the value-betting regime (veto bypass + score floor).
+        is_value_tier = tier_value in self.VALUE_TIER_NAMES
+
         if divergence is not None:
             if divergence >= self.HARD_DIVERGENCE and not is_value_sniper:
                 score -= 42.0
@@ -254,22 +601,10 @@ class BettingBrain:
                 score += 11.0
                 positives.append("v25_v27_aligned")
 
-        # Trap market detection: market overpriced vs historical band hit rate
-        trap_market_flag = False
-        trap_market_gap = None
-        if isinstance(triple, dict):
-            band_rate_val = self._safe_float(triple.get("band_rate"))
-            implied_val = self._safe_float(triple.get("implied_prob"))
-            if (
-                band_rate_val is not None
-                and implied_val is not None
-                and band_sample >= self.MIN_BAND_SAMPLE
-                and (implied_val - band_rate_val) > self.TRAP_MARKET_GAP
-            ):
-                trap_market_flag = True
-                trap_market_gap = round(implied_val - band_rate_val, 4)
-                score -= 14.0
-                issues.append("trap_market_market_overpriced")
+        # Trap market score penalty (flag computed above, before tier check)
+        if trap_market_flag:
+            score -= 14.0
+            issues.append("trap_market_market_overpriced")
 
         if isinstance(triple, dict):
             if triple_is_value:
@@ -371,6 +706,120 @@ class BettingBrain:
             if sniper_bypassed:
                 positives.append("sniper_bypassed_soft_vetoes")
 
+        # ── V31d: VALUE-TIER REGIME ──────────────────────────────────────
+        # A matched MS value tier is the validated profitable signal (60d:
+        # premium 6.0-7.5 → +32.7% ROI). Underdogs are bet on the odds
+        # premium, not on win-probability, so:
+        #   (1) waive the two favorite-confidence vetoes (genuine safety
+        #       vetoes — extreme_neg_ev, ev_too_high_trap, htft_reversal
+        #       _risk_high, v25_v27_hard_disagreement, low_reliability_hard
+        #       — are NOT waived and still reject);
+        #   (2) replace the favorite-confidence SCORE with a value floor so
+        #       premium can clear MIN_BET_SCORE while strong/standard stay
+        #       WATCH-level. All rich analysis output is untouched.
+        value_tier_bypassed: List[str] = []
+        if is_value_tier:
+            if vetoes:
+                remaining = []
+                for v in vetoes:
+                    if v in self.VALUE_TIER_BYPASSABLE_VETOES:
+                        value_tier_bypassed.append(v)
+                    else:
+                        remaining.append(v)
+                vetoes = remaining
+            if value_tier_bypassed:
+                positives.append("value_tier_bypassed_favorite_vetoes")
+            # Value-regime score: floor by tier quality + small +EV nudge.
+            value_score = self.VALUE_TIER_BASE_SCORE.get(tier_value, 50.0)
+            value_score += max(-5.0, min(10.0, ev_edge * 35.0))
+            if odds_rel >= 0.45:
+                value_score += 3.0
+            # Only premium is auto-staked; cap the rest below MIN_BET_SCORE
+            # so they surface as WATCH (visible analysis, not a staked bet).
+            if tier_value != "premium":
+                value_score = min(value_score, 60.0)
+            score = value_score
+            positives.append(f"value_tier_{tier_value}")
+
+        # ── V31f: NATIONAL-TEAM REGIME (overrides club logic) ─────────────
+        # For national matches the validated strategy is a narrow, mechanical
+        # value gate (MS / odds 4-7 / Hazırlık+Eleme). We REPLACE the club
+        # verdict so club-tuned vetoes/scores don't distort it. All the rich
+        # analysis (probs, model_gap, divergence, triple) above is preserved
+        # in the payload below — only action/score/stake are decided here.
+        national_gate_passed = False
+        if is_national:
+            in_band = self.NATIONAL_MIN_ODDS <= odds <= self.NATIONAL_MAX_ODDS
+            is_bet_market = market == self.NATIONAL_BET_MARKET
+            comp_ok = competition_type in self.NATIONAL_ALLOWED_COMPETITIONS
+            # Genuine safety vetoes still kill the bet even for national matches.
+            hard_unsafe = {
+                "low_reliability_league_hard_block",
+                "v25_v27_hard_disagreement",
+                "extreme_negative_ev",
+                "htft_reversal_risk_high",
+            }
+            has_hard_unsafe = any(v in hard_unsafe for v in vetoes)
+
+            national_vetoes: List[str] = []
+            if not is_bet_market:
+                national_vetoes.append("national_non_ms_market_muted")
+            if not comp_ok:
+                national_vetoes.append("national_tournament_no_bet")
+            if not in_band:
+                national_vetoes.append("national_odds_outside_value_band")
+            if has_hard_unsafe:
+                national_vetoes.append("national_hard_safety_veto")
+
+            if national_vetoes:
+                vetoes = national_vetoes
+                action = "REJECT"
+                score = min(score, 40.0)
+                issues.append(f"national_gate:{competition_type or 'unknown'}")
+            else:
+                vetoes = []
+                national_gate_passed = True
+                score = self.NATIONAL_BASE_SCORE + max(-4.0, min(8.0, ev_edge * 30.0))
+                score = max(0.0, min(100.0, score))
+                action = "BET"
+                positives.append("national_value_gate_passed")
+                issues.append(f"national_gate:{competition_type}")
+            # skip the club action logic below
+            row["betting_brain"] = {
+                "action": action,
+                "score": round(score, 1),
+                "summary": self._summary(action, market, pick, positives, issues, vetoes),
+                "positives": positives[:5],
+                "issues": issues[:6],
+                "vetoes": vetoes[:6],
+                "sniper_bypassed": sniper_bypassed,
+                "value_tier_bypassed": [],
+                "is_value_tier": False,
+                "is_national": True,  # V31f
+                "competition_type": competition_type,
+                "trap_market_flag": trap_market_flag,
+                "trap_market_gap": trap_market_gap,
+                "tier_label": "national_value" if national_gate_passed else None,
+                "value_tier": None,
+                "model_prob": round(model_prob, 4) if model_prob is not None else None,
+                "implied_prob": round(implied, 4),
+                "model_market_gap": round(model_gap, 4) if model_gap is not None else None,
+                "v27_prob": round(v27_prob, 4) if v27_prob is not None else None,
+                "divergence": round(divergence, 4) if divergence is not None else None,
+                "triple_key": triple_key,
+                "triple_value": triple,
+            }
+            if national_gate_passed:
+                row["is_guaranteed"] = False  # high-variance band, never "guaranteed"
+                row["pick_reason"] = "national_value_gate"
+                row["stake_units"] = self.NATIONAL_STAKE_UNITS
+                row["bet_grade"] = "B"
+                row["playable"] = True
+            else:
+                self._force_no_bet(row, f"betting_brain_{action.lower()}")
+            self._append_reason(row, f"betting_brain_national_{action.lower()}_{round(score)}")
+            return row
+
         score = max(0.0, min(100.0, score))
         action = "BET"
         if vetoes:
@@ -393,8 +842,12 @@ class BettingBrain:
             "issues": issues[:6],
             "vetoes": vetoes[:6],
             "sniper_bypassed": sniper_bypassed,
+            "value_tier_bypassed": value_tier_bypassed,  # V31d
+            "is_value_tier": is_value_tier,  # V31d
             "trap_market_flag": trap_market_flag,
             "trap_market_gap": trap_market_gap,
+            "tier_label": tier_label,
+            "value_tier": tier_value,  # V31c: premium/strong/standard
             "model_prob": round(model_prob, 4) if model_prob is not None else None,
             "implied_prob": round(implied, 4),
             "model_market_gap": round(model_gap, 4) if model_gap is not None else None,
@@ -407,10 +860,15 @@ class BettingBrain:
         if action != "BET":
             self._force_no_bet(row, f"betting_brain_{action.lower()}")
         else:
-            row["is_guaranteed"] = bool(score >= 82.0)
+            # V31d: value-tier underdogs are high-variance (~20% hit) — never
+            # label them "guaranteed" no matter how high the value score is.
+            row["is_guaranteed"] = bool(score >= 82.0) and not is_value_tier
             row["pick_reason"] = "betting_brain_approved"
             row["stake_units"] = self._brain_stake(row, score)
-            row["bet_grade"] = "A" if score >= 82.0 else "B"
+            # V31c: bet_grade now reflects value_tier so the UI can show
+            # the bettor which quality band a pick belongs to.
+            row["bet_grade"] = self._grade_from_tier(tier_value, score)
+            row["value_tier"] = tier_value
             row["playable"] = True
 
         self._append_reason(row, f"betting_brain_{action.lower()}_{round(score)}")
@@ -465,9 +923,13 @@ class BettingBrain:
             prob = self._safe_float(probs.get(pick), 0.0)
             if prob is None or prob <= 0.0:
                 continue
-            implied_odd = round(1.0 / prob, 2) if prob > 0.01 else 0.0
-            ref_odd = existing_odds_by_pick.get(pick) or implied_odd
-            rows[key] = {
+            # V35c: only REAL bookmaker odds may be displayed. The old fallback
+            # showed synthetic fair-odds (1/prob) as "Oran" — users could read
+            # it as an offered price (e.g. X shown at 4.53 while the bulletin
+            # offered ~3.58). No real price → odds 0.0 and the FE renders "-".
+            real = self._safe_float(getattr(self, "_ms_real_odds", {}).get(pick), 0.0) or 0.0
+            ref_odd = existing_odds_by_pick.get(pick) or (real if real > 1.01 else 0.0)
+            row = {
                 "market": "MS",
                 "pick": pick,
                 "probability": round(prob, 4),
@@ -481,6 +943,12 @@ class BettingBrain:
                 "bet_grade": "PASS",
                 "decision_reasons": ["underdog_reference_for_completeness"],
             }
+            if ref_odd > 1.01:
+                # honest economics vs the real price (vig shows as it truly is)
+                row["implied_prob"] = round(1.0 / ref_odd, 4)
+                row["ev_edge"] = round(prob * ref_odd - 1.0, 4)
+                row["edge"] = row["ev_edge"]
+            rows[key] = row
 
     @staticmethod
     def _merge_row(existing: Optional[Dict[str, Any]], incoming: Dict[str, Any]) -> Dict[str, Any]:
@@ -506,12 +974,14 @@ class BettingBrain:
 
     def _summary_item(self, row: Dict[str, Any]) -> Dict[str, Any]:
         reasons = list(row.get("decision_reasons") or row.get("reasons") or [])
+        brain = row.get("betting_brain") or {}
         return {
             "market": row.get("market"),
             "pick": row.get("pick"),
             "raw_confidence": row.get("raw_confidence", row.get("confidence")),
             "calibrated_confidence": row.get("calibrated_confidence", row.get("confidence")),
             "bet_grade": row.get("bet_grade", "PASS"),
+            "value_tier": row.get("value_tier") or brain.get("value_tier"),  # V31c
             "playable": bool(row.get("playable")),
             "stake_units": float(row.get("stake_units", 0.0) or 0.0),
             "play_score": row.get("play_score", 0.0),
@@ -521,9 +991,22 @@ class BettingBrain:
             "odds": row.get("odds", 0.0),
             "reasons": reasons[:6],
             "is_underdog_reference": bool(row.get("is_underdog_reference")),
-            "betting_brain": row.get("betting_brain"),
+            "betting_brain": brain,
         }
 
+    @staticmethod
+    def _grade_from_tier(value_tier: Optional[str], score: float) -> str:
+        """V31c: Map value_tier → bet grade so the UI surfaces the
+        quality band. Falls back to score-based grade for untiered picks.
+            premium  → A  (deep underdog, highest ROI, high variance)
+            strong   → B  (strong underdog)
+            standard → C  (volume zone, thin edge)
+        """
+        mapping = {"premium": "A", "strong": "B", "standard": "C"}
+        if value_tier in mapping:
+            return mapping[value_tier]
+        return "A" if score >= 82.0 else "B"
+
     @staticmethod
     def _candidate_sort_key(row: Dict[str, Any]) -> Tuple[float, float, float]:
         brain = row.get("betting_brain") or {}
@@ -563,6 +1046,12 @@ class BettingBrain:
         odds = self._safe_float(row.get("odds"), 0.0) or 0.0
         if odds <= 1.0:
             return 0.0
+        # V31d: value-tier underdogs use a small FLAT stake (high variance),
+        # not the score-scaled favorite stake. score is high (70+) by design
+        # but that reflects validated tier EV, not win-probability.
+        brain = row.get("betting_brain") or {}
+        if brain.get("is_value_tier") or brain.get("value_tier") in self.VALUE_TIER_NAMES:
+            return self.VALUE_TIER_STAKE_UNITS
         cap = 2.0 if score >= 82.0 else 1.2
         if score < 78.0:
             cap = 0.8
@@ -635,6 +1124,112 @@ class BettingBrain:
             return self._safe_float(ou25.get(key)) if key else None
         return None
 
+    def _score_consistent_markets(self, package: Dict[str, Any]) -> Optional[set]:
+        """Build the set of (market, pick) tuples that WOULD WIN if the
+        model's own score prediction came true. We use this as a coherence
+        gate: if the model is confident about a 1-0 outcome but also wants
+        to play OU25 Üst, those two beliefs contradict each other — and the
+        score prediction is the more informative one because it aggregates
+        all market signals into a single most-likely scenario.
+
+        Returns None if the score prediction is missing or malformed; in
+        that case we skip the coherence check.
+        """
+        score_pred = package.get("score_prediction") or {}
+        ft_raw = str(score_pred.get("ft") or score_pred.get("full_time") or "").strip()
+        ht_raw = str(score_pred.get("ht") or score_pred.get("half_time") or "").strip()
+
+        def parse(s: str) -> Optional[Tuple[int, int]]:
+            for sep in ("-", ":", "–"):
+                if sep in s:
+                    parts = s.split(sep, 1)
+                    try:
+                        return int(parts[0].strip()), int(parts[1].strip())
+                    except (ValueError, IndexError):
+                        return None
+            return None
+
+        ft = parse(ft_raw)
+        if ft is None:
+            return None
+        ht = parse(ht_raw)
+
+        fh, fa = ft
+        total = fh + fa
+        consistent: set = set()
+
+        # MS / 1X2 — single outcome
+        if fh > fa:
+            consistent.add(("MS", "1"))
+            consistent.add(("ML", "1"))
+        elif fh < fa:
+            consistent.add(("MS", "2"))
+            consistent.add(("ML", "2"))
+        else:
+            consistent.add(("MS", "X"))
+            consistent.add(("ML", "X"))
+
+        # DC — two of three legs win at any score
+        if fh >= fa:
+            consistent.add(("DC", "1X"))
+        if fh <= fa:
+            consistent.add(("DC", "X2"))
+        if fh != fa:
+            consistent.add(("DC", "12"))
+
+        # Over/Under main lines
+        for line, market in ((0.5, "OU05"), (1.5, "OU15"),
+                              (2.5, "OU25"), (3.5, "OU35"), (4.5, "OU45")):
+            if total > line:
+                for p in ("Üst", "Ust", "Over", "OVER"):
+                    consistent.add((market, p))
+            elif total < line:
+                for p in ("Alt", "Under", "UNDER"):
+                    consistent.add((market, p))
+            # total == line → push, neither side wins → don't add
+
+        # BTTS — both teams score
+        if fh > 0 and fa > 0:
+            for p in ("Var", "KG Var", "Yes", "YES"):
+                consistent.add(("BTTS", p))
+        else:
+            for p in ("Yok", "KG Yok", "No", "NO"):
+                consistent.add(("BTTS", p))
+
+        # OE — total goals odd/even
+        if total % 2 == 1:
+            for p in ("Tek", "Odd", "ODD"):
+                consistent.add(("OE", p))
+        else:
+            for p in ("Çift", "Cift", "Even", "EVEN"):
+                consistent.add(("OE", p))
+
+        # HT-only markets (need HT score)
+        if ht is not None:
+            hh, ha = ht
+            ht_total = hh + ha
+            if hh > ha:
+                consistent.add(("HT", "1"))
+            elif hh < ha:
+                consistent.add(("HT", "2"))
+            else:
+                consistent.add(("HT", "X"))
+            for line, market in ((0.5, "HT_OU05"), (1.5, "HT_OU15"), (2.5, "HT_OU25")):
+                if ht_total > line:
+                    for p in ("Üst", "Ust", "Over"):
+                        consistent.add((market, p))
+                elif ht_total < line:
+                    for p in ("Alt", "Under"):
+                        consistent.add((market, p))
+
+            # HTFT — single combo
+            ht_o = "1" if hh > ha else "2" if hh < ha else "X"
+            ft_o = "1" if fh > fa else "2" if fh < fa else "X"
+            consistent.add(("HTFT", f"{ht_o}/{ft_o}"))
+            consistent.add(("HTFT", f"{ht_o}{ft_o}"))
+
+        return consistent
+
     def _triple_value(self, package: Dict[str, Any], key: Optional[str]) -> Optional[Dict[str, Any]]:
         if not key:
             return None

ai-engine/services/orchestrator/data_loader.py

@@ -449,6 +449,12 @@ class DataLoaderMixin:
             return 1.5, 1.2
         return weighted_for / total_weight, weighted_against / total_weight
 
+    # Approximate European season window — Eredivisie/PL/La Liga start late
+    # July / mid-August, end May. Using 300 days as a buffer covers most
+    # competitions while excluding "career points" from previous seasons.
+    # When a proper seasons table lands this should query season boundaries.
+    _SEASON_LOOKBACK_MS = 300 * 24 * 60 * 60 * 1000
+
     def _estimate_league_position(
         self,
         cur: RealDictCursor,
@@ -458,6 +464,7 @@ class DataLoaderMixin:
     ) -> int:
         if not team_id or not league_id:
             return 10
+        season_start_ms = before_date_ms - self._SEASON_LOOKBACK_MS
         try:
             cur.execute(
                 """
@@ -478,6 +485,7 @@ class DataLoaderMixin:
                       AND m.score_home IS NOT NULL
                       AND m.score_away IS NOT NULL
                       AND m.mst_utc < %s
+                      AND m.mst_utc >= %s
                     UNION ALL
                     SELECT
                         m.away_team_id AS team_id,
@@ -492,11 +500,15 @@ class DataLoaderMixin:
                       AND m.score_home IS NOT NULL
                       AND m.score_away IS NOT NULL
                       AND m.mst_utc < %s
+                      AND m.mst_utc >= %s
                 ) tm
                 GROUP BY tm.team_id
                 ORDER BY points DESC
                 """,
-                (league_id, before_date_ms, league_id, before_date_ms),
+                (
+                    league_id, before_date_ms, season_start_ms,
+                    league_id, before_date_ms, season_start_ms,
+                ),
             )
             rows = cur.fetchall()
             if not rows:

ai-engine/services/orchestrator/feature_builder.py

@@ -225,20 +225,43 @@ class FeatureBuilderMixin:
         if enrichment_failures:
             print(f"⚠️ Enrichment partial failures for {data.match_id}: {', '.join(enrichment_failures)}")
 
+        # ── Cup game detection (used by upset engine + elo dampening below) ──
+        _league_name_lower = (getattr(data, 'league_name', '') or '').lower()
+        _cup_keywords = ("kupa", "cup", "coupe", "copa", "coppa", "pokal",
+                         "trophy", "shield", "ziraat", "süper kupa", "super cup",
+                         "beker", "taça", "taca")
+        _is_cup_match = any(kw in _league_name_lower for kw in _cup_keywords)
+
+        # ── League size hint: top European leagues 18-20 teams, lower 16-24 ──
+        # We don't have a per-league team count, so fall back to 20 (standard).
+        # When standings infra lands this should pull from seasons table.
+        _league_total_teams = 20
+
         # Upset engine features
         upset_atmosphere, upset_motivation, upset_fatigue = 0.0, 0.0, 0.0
         try:
             upset_engine = get_upset_engine()
+            # Use the real position estimates from data_loader; fall back to mid-
+            # table (10) only when the loader couldn't compute one. Hardcoding 10
+            # for every team made motivation_score collapse to 0 for everyone.
+            _home_pos = getattr(data, 'home_position', None)
+            _away_pos = getattr(data, 'away_position', None)
+            if _home_pos is None or _home_pos <= 0:
+                _home_pos = 10
+            if _away_pos is None or _away_pos <= 0:
+                _away_pos = 10
             upset_feats = upset_engine.get_features(
                 home_team_name=getattr(data, 'home_team_name', '') or '',
                 home_team_id=data.home_team_id,
                 away_team_name=getattr(data, 'away_team_name', '') or '',
                 league_name=getattr(data, 'league_name', '') or '',
-                home_position=10,
-                away_position=10,
+                home_position=_home_pos,
+                away_position=_away_pos,
                 match_date_ms=data.match_date_ms,
+                is_cup_match=_is_cup_match,
                 home_days_rest=int(home_rest),
                 away_days_rest=int(away_rest),
+                total_teams=_league_total_teams,
             )
             upset_atmosphere = upset_feats.get('upset_atmosphere', 0.0)
             upset_motivation = upset_feats.get('upset_motivation', 0.0)
@@ -276,15 +299,10 @@ class FeatureBuilderMixin:
         is_season_start = 1.0 if match_month in (7, 8, 9) else 0.0
         is_season_end = 1.0 if match_month in (5, 6) else 0.0
 
-        # ── Cup game detection: dampen home advantage in feature space ──
-        _league_name = (getattr(data, 'league_name', '') or '').lower()
-        _cup_keywords = ("kupa", "cup", "coupe", "copa", "coppa", "pokal",
-                         "trophy", "shield", "ziraat", "süper kupa", "super cup")
-        _is_cup = any(kw in _league_name for kw in _cup_keywords)
-
         # ── Derived / Interaction features (V27) ──
         # Cup games: home ELO advantage is ~30% weaker (rotation, lower motivation)
-        elo_diff = (home_elo - away_elo) * (0.70 if _is_cup else 1.0)
+        # Uses _is_cup_match computed earlier (before upset engine call).
+        elo_diff = (home_elo - away_elo) * (0.70 if _is_cup_match else 1.0)
         form_elo_diff = home_form_elo_val - away_form_elo_val
         attack_vs_defense_home = data.home_goals_avg - data.away_conceded_avg
         attack_vs_defense_away = data.away_goals_avg - data.home_conceded_avg

ai-engine/services/orchestrator/market_board.py

@@ -56,10 +56,81 @@ from services.match_commentary import generate_match_commentary
 from utils.top_leagues import load_top_league_ids
 from utils.league_reliability import load_league_reliability
 from config.config_loader import build_threshold_dict, get_threshold_default
-from models.calibration import get_calibrator
+from models.calibration import get_calibrator, get_final_recalibrator
+from models.market_anchor import devig, apply_corrections
+from models.score_matrix import build_calibrated_score_package
+from models.live_matrix import build_live_projection, estimate_minute
+
+# ── V30: Post-calibration trust factors ─────────────────────────────
+# Controls how much to trust isotonic calibrator vs raw model output.
+# trust=1.0 → use calibrator fully; trust=0.0 → bypass, use raw model.
+# Derived from calibrator_metrics.json analysis (mean_predicted vs mean_actual):
+#   MS calibrators: gap < 0.5%  → excellent, full trust
+#   BTTS:           gap = +14.4% → calibrator broken, bypass
+#   OU25:           gap = +5.3%  → over-inflates, mostly bypass
+#   OU35:           gap = +3.6%  → moderate inflation, dampen
+#   OU15:           gap = +1.5%  → slight, mostly trust
+#   HT:             mixed        → moderate trust
+#   DC/HT_FT:       < 30 samples → unreliable, bypass
+POST_CAL_TRUST: Dict[str, float] = {
+    "ms_home": 1.0,
+    "ms_draw": 1.0,
+    "ms_away": 1.0,
+    "btts":    0.0,
+    "ou25":    0.15,
+    "ou35":    0.30,
+    "ou15":    0.70,
+    "ht_home": 0.50,
+    "ht_draw": 0.30,
+    "ht_away": 0.50,
+    "dc":      0.0,
+    "ht_ft":   0.0,
+}
 
 
 class MarketBoardMixin:
+    def _league_confidence_for(self, league_id: Optional[str]) -> Optional[Dict[str, Any]]:
+        """Return the backtest-derived confidence record for a league, or None.
+
+        Shape: {"label": high|medium|low, "bet_roi": float, "bet_n": int,
+        "hit": float}. None → league absent or too few bets ('unknown') → FE
+        shows no badge. Never raises (missing artifact = graceful None)."""
+        if not league_id:
+            return None
+        lookup = getattr(self, "league_confidence", None) or {}
+        info = lookup.get(str(league_id))
+        if not isinstance(info, dict):
+            return None
+        label = info.get("label")
+        if label in (None, "unknown"):
+            return None
+        return {
+            "label": label,
+            "bet_roi": info.get("bet_roi"),
+            "bet_n": info.get("bet_n"),
+            "hit": info.get("hit"),
+        }
+
+    def _is_national_match(self, league_id: Optional[str]) -> bool:
+        """True if this league is an A-milli (senior men's) national competition."""
+        if not league_id:
+            return False
+        natl = getattr(self, "national_leagues", None) or set()
+        return str(league_id) in natl
+
+    def _competition_type_for(
+        self, league_id: Optional[str], league_name: Optional[str]
+    ) -> Optional[str]:
+        """For national matches, classify HAZIRLIK/ELEME/TURNUVA from the league
+        name. None for non-national leagues (clubs don't use this)."""
+        if not self._is_national_match(league_id):
+            return None
+        try:
+            from utils.national_leagues import classify_competition
+            return classify_competition(league_name or "")
+        except Exception:
+            return None
+
     def _build_prediction_package(
         self,
         data: MatchData,
@@ -280,6 +351,34 @@ class MarketBoardMixin:
             if market in available_markets
         }
 
+        # V35: anchor the DISPLAYED per-market probabilities to the de-vigged
+        # market price (+ proven home-favourite correction). The model's own
+        # numbers were measured ~25-30% mis-calibrated; the de-vigged market is
+        # ~1.5% (out-of-sample). This only rewrites what the user sees.
+        market_board = self._apply_market_anchor(market_board, data)
+
+        # V35b: make the DISPLAYED confidence/edge fields on every pick object
+        # consistent with the calibrated board (Güven Skoru, Güven Aralığı,
+        # Model%/Teorik-avantaj), then drop a "value pick" that has no real edge
+        # once priced honestly — no fabricated value bets.
+        self._apply_anchor_to_picks(
+            market_board, main_pick, value_pick, aggressive_pick, supporting, bet_summary,
+        )
+        if value_pick is not None and float(value_pick.get("ev_edge", 0.0) or 0.0) <= 0.0:
+            value_pick = None
+
+        # V36: derive the score card (score_prediction + scenario_top5) from the
+        # SAME anchored probabilities, so it can never contradict the MS card.
+        # Validated on 63,681 real-odds matches: modal-score hit 12.6% vs stated
+        # 13.1%, top-5 coverage 51%, per-score gaps <1.2pt.
+        cal_score = self._build_calibrated_score(market_board)
+
+        # V38: while the match is LIVE, also project score/minute-conditioned
+        # probabilities (P(side scores again), live 1X2, comeback, scenarios).
+        # OOS-validated on 70,410 reconstructed live moments: ECE 0.5-0.8%;
+        # "one-goal lead at 80'" case: said 21.7% vs actual 23.0%.
+        live_projection = self._build_live_projection(market_board, data)
+
         # Determine simulation mode for the response
         _resp_status = str(data.status or "").upper()
         _resp_state  = str(data.state  or "").upper()
@@ -294,6 +393,13 @@ class MarketBoardMixin:
                 "home_team": data.home_team_name,
                 "away_team": data.away_team_name,
                 "league": data.league_name,
+                "league_id": data.league_id,
+                # Backtest-derived per-league confidence (ROI + sample size).
+                # None when the league has too little data to judge → FE shows no badge.
+                "league_confidence": self._league_confidence_for(data.league_id),
+                # National-team match flags (drive betting_brain's national gate).
+                "is_national": self._is_national_match(data.league_id),
+                "competition_type": self._competition_type_for(data.league_id, data.league_name),
                 "match_date_ms": data.match_date_ms,
                 "sport": data.sport,
                 # Live snapshot — match_commentary uses this to detect upset-in-progress
@@ -332,15 +438,24 @@ class MarketBoardMixin:
             "bet_summary": bet_summary,
             "supporting_picks": supporting,
             "aggressive_pick": aggressive_pick,
-            "scenario_top5": prediction.ft_scores_top5,
-            "score_prediction": {
-                "ft": prediction.predicted_ft_score,
-                "ht": prediction.predicted_ht_score,
-                "xg_home": round(float(prediction.home_xg), 2),
-                "xg_away": round(float(prediction.away_xg), 2),
-                "xg_total": round(float(prediction.total_xg), 2),
-            },
+            "scenario_top5": (
+                cal_score["scenario_top5"] if cal_score else prediction.ft_scores_top5
+            ),
+            "score_prediction": (
+                cal_score["score_prediction"]
+                if cal_score
+                else {
+                    "ft": prediction.predicted_ft_score,
+                    "ht": prediction.predicted_ht_score,
+                    "xg_home": round(float(prediction.home_xg), 2),
+                    "xg_away": round(float(prediction.away_xg), 2),
+                    "xg_total": round(float(prediction.total_xg), 2),
+                }
+            ),
             "market_board": market_board,
+            # V38: score/minute-aware live probabilities (None when not live or
+            # no real odds). FE can render "deplasman gol atar: %X / dönme: %Y".
+            "live_projection": live_projection,
             "others": {
                 "handicap": prediction.handicap_pick,
                 "cards": {
@@ -942,6 +1057,301 @@ class MarketBoardMixin:
             }
         return merged
 
+    # ── V35 market-anchored calibration ────────────────────────────────
+    # Maps a board pick label -> the probs key it refers to, so the displayed
+    # confidence can be set to the EXISTING pick's now-calibrated probability.
+    _ANCHOR_PICK_KEY: Dict[str, Dict[str, str]] = {
+        "MS":      {"1": "1", "X": "X", "0": "X", "2": "2"},
+        "HT":      {"1": "1", "X": "X", "0": "X", "2": "2"},
+        "DC":      {"1X": "1X", "X2": "X2", "12": "12",
+                    "1-X": "1X", "X-2": "X2", "1-2": "12"},
+        "OU15":    {"Üst": "over", "Alt": "under", "Over": "over", "Under": "under"},
+        "OU25":    {"Üst": "over", "Alt": "under", "Over": "over", "Under": "under"},
+        "OU35":    {"Üst": "over", "Alt": "under", "Over": "over", "Under": "under"},
+        "HT_OU05": {"Üst": "over", "Alt": "under", "Over": "over", "Under": "under"},
+        "HT_OU15": {"Üst": "over", "Alt": "under", "Over": "over", "Under": "under"},
+        "BTTS":    {"KG Var": "yes", "KG Yok": "no", "Var": "yes", "Yok": "no",
+                    "Yes": "yes", "No": "no"},
+        "OE":      {"Tek": "odd", "Çift": "even", "Odd": "odd", "Even": "even"},
+    }
+
+    def _set_board(
+        self,
+        market_board: Dict[str, Any],
+        market: str,
+        probs: Dict[str, float],
+    ) -> None:
+        """Overwrite one board entry's probs with calibrated values and refresh
+        its confidence to the EXISTING pick's now-calibrated probability.
+
+        We recalibrate the NUMBERS, not the pick selection — showing the engine's
+        pick alongside its honest probability. Falls back to the most-likely
+        outcome only when the pick can't be mapped."""
+        entry = market_board.get(market)
+        if not isinstance(entry, dict):
+            return
+        rounded = {k: round(float(v), 4) for k, v in probs.items()}
+        if not rounded:
+            return
+        entry["probs"] = rounded
+        pick = str(entry.get("pick") or "")
+        key = self._ANCHOR_PICK_KEY.get(market, {}).get(pick)
+        if key is None or key not in rounded:
+            key = max(rounded, key=rounded.get)
+        entry["confidence"] = round(rounded[key] * 100.0, 1)
+        entry["calibration_source"] = "market_anchor_v35"
+
+    def _apply_market_anchor(
+        self,
+        market_board: Dict[str, Any],
+        data: MatchData,
+    ) -> Dict[str, Any]:
+        """Anchor DISPLAYED per-market probabilities to the de-vigged market
+        price (+ proven home-favourite correction for MS, and DC derived from
+        it for internal consistency).
+
+        Only markets with REAL odds are rewritten — `devig` returns None for any
+        missing/placeholder leg, so no-odds markets are left untouched (and are
+        already dropped upstream per the product rule: never show fabricated
+        numbers for a match without odds). Toggle off with env MARKET_ANCHOR_CAL=0.
+        """
+        if os.environ.get("MARKET_ANCHOR_CAL", "1") == "0":
+            return market_board
+        if not isinstance(market_board, dict) or not market_board:
+            return market_board
+        odds = getattr(data, "odds_data", None) or {}
+
+        def real(key: str) -> Optional[float]:
+            val = self._real_market_odds(odds, key)
+            return val if val > 1.01 else None
+
+        # MS (3-way) + favourite corrections; DC derived from the same vector
+        ms = devig([real("ms_h"), real("ms_d"), real("ms_a")])
+        if ms is not None:
+            p1, px, p2 = apply_corrections(*ms)
+            if "MS" in market_board:
+                self._set_board(market_board, "MS", {"1": p1, "X": px, "2": p2})
+            if "DC" in market_board:
+                self._set_board(
+                    market_board, "DC",
+                    {"1X": p1 + px, "X2": px + p2, "12": p1 + p2},
+                )
+
+        # HT (3-way)
+        ht = devig([real("ht_h"), real("ht_d"), real("ht_a")])
+        if ht is not None and "HT" in market_board:
+            self._set_board(market_board, "HT", {"1": ht[0], "X": ht[1], "2": ht[2]})
+
+        # 2-way markets
+        for mk, ko, ku, lo, lu in (
+            ("OU15", "ou15_o", "ou15_u", "over", "under"),
+            ("OU25", "ou25_o", "ou25_u", "over", "under"),
+            ("OU35", "ou35_o", "ou35_u", "over", "under"),
+            ("BTTS", "btts_y", "btts_n", "yes", "no"),
+            ("OE", "oe_odd", "oe_even", "odd", "even"),
+            ("HT_OU05", "ht_ou05_o", "ht_ou05_u", "over", "under"),
+            ("HT_OU15", "ht_ou15_o", "ht_ou15_u", "over", "under"),
+        ):
+            if mk not in market_board:
+                continue
+            pair = devig([real(ko), real(ku)])
+            if pair is not None:
+                self._set_board(market_board, mk, {lo: pair[0], lu: pair[1]})
+
+        return market_board
+
+    def _anchored_prob_for(
+        self,
+        market_board: Dict[str, Any],
+        market: str,
+        pick: Any,
+    ) -> Optional[float]:
+        """Look up a pick's calibrated probability from the anchored board.
+
+        Returns None unless the market was actually anchored (real odds) and the
+        pick maps to a known outcome — so no-odds picks are never touched."""
+        entry = market_board.get(str(market or ""))
+        if not isinstance(entry, dict):
+            return None
+        if entry.get("calibration_source") != "market_anchor_v35":
+            return None
+        probs = entry.get("probs") or {}
+        key = self._ANCHOR_PICK_KEY.get(str(market or ""), {}).get(str(pick or ""))
+        if key is None or key not in probs:
+            return None
+        try:
+            return float(probs[key])
+        except (TypeError, ValueError):
+            return None
+
+    def _recalibrate_pick_display(
+        self,
+        obj: Optional[Dict[str, Any]],
+        market_board: Dict[str, Any],
+    ) -> None:
+        """Rewrite ONE pick object's displayed confidence/edge fields so they are
+        consistent with the calibrated (de-vigged market) probability.
+
+        Fixes Güven Skoru (`calibrated_confidence`/`unified_score`), Güven Aralığı
+        (`confidence_interval` recentred on the calibrated confidence), and the
+        value card's Model%/Teorik-avantaj (`model_probability`/`ev_edge`/`edge`,
+        recomputed honestly against the real price → the vig shows as it truly is,
+        no fabricated positive edge). Selection/gates/stake are left untouched."""
+        if not isinstance(obj, dict):
+            return
+        p = self._anchored_prob_for(market_board, obj.get("market"), obj.get("pick"))
+        if p is None:
+            return
+        try:
+            odds = float(obj.get("odds") or 0.0)
+        except (TypeError, ValueError):
+            odds = 0.0
+        implied = (1.0 / odds) if odds > 1.0 else 0.0
+        conf = round(p * 100.0, 1)
+        ev = round(p * odds - 1.0, 4) if odds > 1.0 else 0.0
+        obj["calibrated_probability"] = round(p, 4)
+        obj["model_probability"] = round(p, 4)
+        obj["calibrated_confidence"] = conf
+        obj["unified_score"] = conf
+        obj["implied_prob"] = round(implied, 4)
+        obj["model_edge"] = round(p - implied, 4) if implied > 0.0 else 0.0
+        obj["ev_edge"] = ev
+        obj["edge"] = ev
+        # Recentre the confidence interval on the calibrated confidence, keeping a
+        # sensible width (preserve the engine's width when present).
+        width = 16.0
+        ci = obj.get("confidence_interval")
+        if isinstance(ci, dict) and ci.get("lower") is not None and ci.get("upper") is not None:
+            try:
+                width = max(6.0, float(ci["upper"]) - float(ci["lower"]))
+            except (TypeError, ValueError):
+                width = 16.0
+        half = width / 2.0
+        lower = round(max(0.0, conf - half), 1)
+        upper = round(min(100.0, conf + half), 1)
+        band = "HIGH" if conf >= 60.0 else "MEDIUM" if conf >= 42.0 else "LOW"
+        obj["confidence_interval"] = {
+            "band": band,
+            "lower": lower,
+            "upper": upper,
+            "width": round(upper - lower, 1),
+            "threshold_met": conf >= 50.0,
+        }
+        obj["confidence_band"] = band
+        obj["calibration_source"] = "market_anchor_v35"
+
+    def _apply_anchor_to_picks(
+        self,
+        market_board: Dict[str, Any],
+        main_pick: Optional[Dict[str, Any]],
+        value_pick: Optional[Dict[str, Any]],
+        aggressive_pick: Optional[Dict[str, Any]],
+        supporting: Optional[List[Dict[str, Any]]],
+        bet_summary: Optional[List[Dict[str, Any]]],
+    ) -> None:
+        """Make every DISPLAYED pick object consistent with the anchored board.
+        Toggle off with env MARKET_ANCHOR_CAL=0."""
+        if os.environ.get("MARKET_ANCHOR_CAL", "1") == "0":
+            return
+        for obj in (main_pick, value_pick, aggressive_pick):
+            self._recalibrate_pick_display(obj, market_board)
+        for obj in list(supporting or []):
+            self._recalibrate_pick_display(obj, market_board)
+        for obj in list(bet_summary or []):
+            self._recalibrate_pick_display(obj, market_board)
+
+    def _build_calibrated_score(
+        self,
+        market_board: Dict[str, Any],
+    ) -> Optional[Dict[str, Any]]:
+        """V36: score card derived from the anchored MS + OU25 probabilities.
+
+        Returns {"score_prediction": {...}, "scenario_top5": [...]} or None when
+        the needed markets weren't anchored (no real odds) — in which case the
+        caller keeps the model's own score output. Same kill-switch as V35."""
+        if os.environ.get("MARKET_ANCHOR_CAL", "1") == "0":
+            return None
+        ms = market_board.get("MS") or {}
+        ou = market_board.get("OU25") or {}
+        if (
+            ms.get("calibration_source") != "market_anchor_v35"
+            or ou.get("calibration_source") != "market_anchor_v35"
+        ):
+            return None
+        try:
+            p1 = float(ms["probs"]["1"])
+            px = float(ms["probs"]["X"])
+            p2 = float(ms["probs"]["2"])
+            p_over = float(ou["probs"]["over"])
+        except (KeyError, TypeError, ValueError):
+            return None
+
+        ht_probs = None
+        ht = market_board.get("HT") or {}
+        if ht.get("calibration_source") == "market_anchor_v35":
+            try:
+                ht_probs = (
+                    float(ht["probs"]["1"]),
+                    float(ht["probs"]["X"]),
+                    float(ht["probs"]["2"]),
+                )
+            except (KeyError, TypeError, ValueError):
+                ht_probs = None
+
+        try:
+            pkg = build_calibrated_score_package(p1, px, p2, p_over, ht_probs=ht_probs)
+        except (ValueError, ZeroDivisionError, OverflowError):
+            return None
+        return {
+            "score_prediction": {
+                "ft": pkg["ft"],
+                "ht": pkg["ht"],
+                "xg_home": pkg["xg_home"],
+                "xg_away": pkg["xg_away"],
+                "xg_total": pkg["xg_total"],
+                "ht_top3": pkg["ht_top"],
+                "calibration_source": pkg["calibration_source"],
+            },
+            "scenario_top5": pkg["scenario_top5"],
+        }
+
+    def _build_live_projection(
+        self,
+        market_board: Dict[str, Any],
+        data: MatchData,
+    ) -> Optional[Dict[str, Any]]:
+        """V38: score/minute-conditioned live projection from the anchored
+        probabilities. None unless the match is live, both MS and OU25 were
+        anchored (real odds) and a minute estimate exists. Same kill-switch."""
+        if os.environ.get("MARKET_ANCHOR_CAL", "1") == "0":
+            return None
+        if not self._is_live_match(data):
+            return None
+        ms = market_board.get("MS") or {}
+        ou = market_board.get("OU25") or {}
+        if (
+            ms.get("calibration_source") != "market_anchor_v35"
+            or ou.get("calibration_source") != "market_anchor_v35"
+        ):
+            return None
+        minute = estimate_minute(
+            getattr(data, "match_date_ms", None), int(time.time() * 1000)
+        )
+        if minute is None:
+            return None
+        try:
+            return build_live_projection(
+                float(ms["probs"]["1"]),
+                float(ms["probs"]["X"]),
+                float(ms["probs"]["2"]),
+                float(ou["probs"]["over"]),
+                int(data.current_score_home or 0),
+                int(data.current_score_away or 0),
+                minute,
+            )
+        except (KeyError, TypeError, ValueError, ZeroDivisionError, OverflowError):
+            return None
+
     def _build_market_rows(
         self,
         data: MatchData,
@@ -1114,10 +1524,31 @@ class MarketBoardMixin:
         if cal_key and cal_key in calibrator.calibrators:
             cal_input = max(0.001, min(0.999, raw_conf / 100.0))
             cal_prob = calibrator.calibrate(cal_key, cal_input, odds_val=odd if odd > 1.0 else None)
+            # V30: Trust-based blending — some calibrators inflate probabilities.
+            # Blend isotonic output with raw model based on calibrator accuracy.
+            trust = POST_CAL_TRUST.get(cal_key, 0.5)
+            cal_prob = trust * cal_prob + (1.0 - trust) * cal_input
             calibrated_conf = max(1.0, min(99.0, cal_prob * 100.0))
         else:
-            multiplier = self.market_calibration.get(market, 0.85)
-            calibrated_conf = max(1.0, min(99.0, raw_conf * multiplier))
+            # V31b: Fallback for markets WITHOUT isotonic calibrator.
+            # Old approach used aggressive multipliers (0.58-0.85) causing
+            # massive deflation: HT_OU15 -40.5%, HT_OU05 -25.2%, OE -18.3%.
+            # New approach: mild damping (0.92) acknowledges slight model
+            # overconfidence without destroying probability signal.
+            # The tier system (V31b) is the real profitability gatekeeper.
+            calibrated_conf = max(1.0, min(99.0, raw_conf * 0.92))
+
+        # ── FINAL-OUTPUT RECALIBRATION (V31e) ──────────────────────────
+        # Last-step per-market map: "system says X% -> reality is Y%". ONLY
+        # badly-miscalibrated markets carry a map (fit-ECE >= 5: OU15, OU35,
+        # HT_OU05, HT_OU15). MS and every already-good market pass through
+        # UNCHANGED -> guaranteed no regression. Out-of-sample proven (e.g.
+        # HT_OU15 ECE 29.2->0.8) and identity-safe for MS (1.1->1.3).
+        # This adjusts ONLY the displayed confidence so users see honest
+        # probabilities; all analysis below (probabilities, edges, vetoes,
+        # tiers, bands) is preserved, and the pre-recal value is kept for audit.
+        pre_recal_conf = calibrated_conf
+        calibrated_conf = get_final_recalibrator().recalibrate_conf(market, calibrated_conf)
         min_conf = self.market_min_conf.get(market, 55.0)
 
         implied_prob = (1.0 / odd) if odd > 1.0 else 0.0
@@ -1178,9 +1609,11 @@ class MarketBoardMixin:
         reasons: List[str] = []
         playable = True
         
-        # V34: Broadened value_sniper bypass — odds-aware model rarely shows 3% EV edge
-        # Allow high-confidence predictions OR modest positive EV to bypass secondary gates
-        is_value_sniper = ev_edge >= 0.008 or calibrated_conf >= 55.0
+        # V29b: Permissive upstream — let betting_brain's tiered system do the real filtering.
+        # Old threshold (ev>=0.008 OR conf>=55) let everything through AND bypassed brain vetoes.
+        # New approach: let most picks through market_board, but brain's MARKET_ODDS_TIERS
+        # + hard vetoes (neg EV, muted, low reliability) handle the intelligent filtering.
+        is_value_sniper = calibrated_conf >= 45.0
 
         if calibrated_conf < min_conf:
             if not is_value_sniper:
@@ -1283,11 +1716,50 @@ class MarketBoardMixin:
             stake_units = 0.25  # minimum stake (conservative)
             reasons.append("no_ev_edge_minimum_stake")
 
+        # ── V30: Birleşik Güven Skoru (BGS) ────────────────────────────
+        # A single, honest metric for users: quality-adjusted win probability.
+        # Combines calibrated probability with data quality signals.
+        # Correlation analysis: model_gap r=-0.12, trap negative, reliability weak positive.
+        bgs = calibrated_conf  # POST_CAL_TRUST corrected base
+        model_gap = prob - implied_prob if implied_prob > 0 else 0.0
+        # Penalty when model overestimates vs market (r=-0.12 correlation)
+        if model_gap > 0.05:
+            bgs -= 8.0
+        elif model_gap > 0.0:
+            bgs -= 3.0
+        # Trap market detection: implied prob significantly above historical band rate
+        is_trap_signal = False
+        if band_available and band_prob > 0 and implied_prob > 0:
+            is_trap_signal = (implied_prob - band_prob) > 0.10
+        if is_trap_signal:
+            bgs -= 7.0
+        # League reliability adjustment (±2)
+        bgs += (odds_rel - 0.50) * 4.0
+        # Band alignment
+        if band_available:
+            if bool(band_verdict.get("aligned")):
+                bgs += 2.0
+            else:
+                bgs -= 3.0
+        # BGS label for frontend
+        bgs = max(1.0, min(99.0, bgs))
+        if bgs >= 70:
+            bgs_label = "very_reliable"
+        elif bgs >= 55:
+            bgs_label = "reliable"
+        elif bgs >= 40:
+            bgs_label = "moderate"
+        else:
+            bgs_label = "low"
+
         out = dict(row)
         out.update(
             {
                 "raw_confidence": round(raw_conf, 1),
                 "calibrated_confidence": round(calibrated_conf, 1),
+                "calibrated_confidence_pre_recal": round(pre_recal_conf, 1),
+                "unified_score": round(bgs, 1),
+                "unified_score_label": bgs_label,
                 "min_required_confidence": round(min_conf, 1),
                 "min_required_play_score": round(min_play_score, 1),
                 "min_required_edge": round(min_edge, 4),
@@ -1347,6 +1819,8 @@ class MarketBoardMixin:
             "pick": row.get("pick"),
             "raw_confidence": row.get("raw_confidence", row.get("confidence")),
             "calibrated_confidence": row.get("calibrated_confidence", row.get("confidence")),
+            "unified_score": row.get("unified_score", row.get("calibrated_confidence", 0.0)),
+            "unified_score_label": row.get("unified_score_label", "moderate"),
             "bet_grade": row.get("bet_grade", "PASS"),
             "playable": bool(row.get("playable")),
             "stake_units": float(row.get("stake_units", 0.0)),

ai-engine/services/orchestrator/upper_brain.py

@@ -60,8 +60,23 @@ from models.calibration import get_calibrator
 
 
 class UpperBrainMixin:
-    def _apply_upper_brain_guards(self, package: Dict[str, Any]) -> Dict[str, Any]:
-        return BettingBrain().judge(package)
+    def _apply_upper_brain_guards(
+        self, package: Dict[str, Any], data: Any = None
+    ) -> Dict[str, Any]:
+        # V35c: hand the brain the REAL bookmaker MS odds so its reference rows
+        # can never display synthetic 1/p prices as if they were offered.
+        ms_real_odds = None
+        if data is not None:
+            try:
+                odds = getattr(data, "odds_data", None) or {}
+                ms_real_odds = {
+                    "1": self._real_market_odds(odds, "ms_h"),
+                    "X": self._real_market_odds(odds, "ms_d"),
+                    "2": self._real_market_odds(odds, "ms_a"),
+                }
+            except Exception:
+                ms_real_odds = None
+        return BettingBrain().judge(package, ms_real_odds=ms_real_odds)
 
         v27_engine = package.get("v27_engine")
         if not isinstance(v27_engine, dict) or not v27_engine.get("triple_value"):

ai-engine/services/single_match_orchestrator.py

@@ -57,6 +57,8 @@ from services.v26_shadow_engine import V26ShadowEngine, get_v26_shadow_engine
 from services.match_commentary import generate_match_commentary
 from utils.top_leagues import load_top_league_ids
 from utils.league_reliability import load_league_reliability
+from utils.league_confidence import load_league_confidence
+from utils.national_leagues import load_national_leagues
 from config.config_loader import build_threshold_dict, get_threshold_default, get_config
 from models.calibration import get_calibrator
 
@@ -171,6 +173,8 @@ class SingleMatchOrchestrator(
         self.engine_mode = str(os.getenv("AI_ENGINE_MODE", "v28-pro-max")).strip().lower()
         self.top_league_ids = load_top_league_ids()
         self.league_reliability = load_league_reliability()
+        self.league_confidence = load_league_confidence()
+        self.national_leagues = load_national_leagues()
         self.enrichment = FeatureEnrichmentService()
         self.odds_band_analyzer = OddsBandAnalyzer()
         # ── Market Thresholds (loaded from config/market_thresholds.json) ──
@@ -664,7 +668,28 @@ class SingleMatchOrchestrator(
                 base_package.setdefault("analysis_details", {})
                 base_package["analysis_details"]["v27_loaded"] = False
 
-        base_package = self._apply_upper_brain_guards(base_package)
+        base_package = self._apply_upper_brain_guards(base_package, data)
+
+        # V35c: the brain rebuilt main/value/supporting/bet_summary AFTER the
+        # market anchor ran inside _build_prediction_package — re-stamp the
+        # calibrated display fields (Güven/CI/Model%/edge) so they stay
+        # consistent, BEFORE the commentary reads the package.
+        self._apply_anchor_to_picks(
+            base_package.get("market_board") or {},
+            base_package.get("main_pick"),
+            base_package.get("value_pick"),
+            base_package.get("aggressive_pick"),
+            base_package.get("supporting_picks"),
+            base_package.get("bet_summary"),
+        )
+        _mp = base_package.get("main_pick")
+        _advice = base_package.get("bet_advice")
+        if isinstance(_mp, dict) and isinstance(_advice, dict) and _mp.get("confidence_band"):
+            _advice["confidence_band"] = _mp["confidence_band"]
+        # no fabricated value bets: a value pick must carry measured positive edge
+        _vp = base_package.get("value_pick")
+        if isinstance(_vp, dict) and float(_vp.get("ev_edge", 0.0) or 0.0) <= 0.0:
+            base_package["value_pick"] = None
 
         # ── Match Commentary: human-readable summary ──────────────
         try:

ai-engine/tests/test_live_matrix.py

@@ -0,0 +1,80 @@
+"""Unit tests for V38 live-conditioned projection (pure, no DB/model deps)."""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+
+from models.live_matrix import (
+    build_live_projection,
+    estimate_minute,
+    state_multiplier,
+)
+
+
+def _approx(a, b, tol=1e-6):
+    return abs(a - b) <= tol
+
+
+def test_probs_form_distribution():
+    proj = build_live_projection(0.50, 0.27, 0.23, 0.55, 1, 0, 60)
+    p = proj["probs"]
+    assert _approx(p["1"] + p["X"] + p["2"], 1.0, 1e-3)
+    assert 0.0 <= proj["p_away_scores_again"] <= 1.0
+
+
+def test_minute_one_roughly_matches_prematch():
+    # at 0-0 minute 1 the projection must stay close to the anchored numbers
+    proj = build_live_projection(0.50, 0.27, 0.23, 0.55, 0, 0, 1)
+    assert abs(proj["probs"]["1"] - 0.50) < 0.06
+    assert abs(proj["probs"]["2"] - 0.23) < 0.06
+
+
+def test_one_goal_lead_at_80():
+    # the user's exact case: 1-0 at 80' (OOS-validated: said 21.7 / actual 23.0)
+    proj = build_live_projection(0.50, 0.27, 0.23, 0.55, 1, 0, 80)
+    assert proj["probs"]["1"] > 0.72                 # leader is now strong fav
+    assert 0.08 <= proj["p_away_scores_again"] <= 0.30
+    assert _approx(
+        proj["p_comeback"], proj["probs"]["X"] + proj["probs"]["2"], 1e-9
+    )
+
+
+def test_less_time_means_fewer_chances():
+    early = build_live_projection(0.50, 0.27, 0.23, 0.55, 1, 0, 60)
+    late = build_live_projection(0.50, 0.27, 0.23, 0.55, 1, 0, 85)
+    assert late["p_away_scores_again"] < early["p_away_scores_again"]
+    assert late["probs"]["1"] > early["probs"]["1"]
+
+
+def test_trailing_team_pushes_late():
+    assert state_multiplier(-1, 80) > 1.05    # trailing by one, late: pushes
+    assert state_multiplier(1, 80) < 1.0      # leading by one, late: parks bus
+    assert state_multiplier(-1, 80) > state_multiplier(-1, 30)
+
+
+def test_score_consistency_with_current_score():
+    proj = build_live_projection(0.50, 0.27, 0.23, 0.55, 2, 1, 75)
+    # every scenario must be reachable from the current score
+    for s in proj["scenario_top5"]:
+        fh, fa = map(int, str(s["score"]).split("-"))
+        assert fh >= 2 and fa >= 1
+    assert proj["current_score"] == "2-1"
+
+
+def test_estimate_minute_approximation():
+    now = 1_700_000_000_000
+    assert estimate_minute(None, now) is None
+    assert estimate_minute(now + 60_000, now) is None          # not kicked off
+    assert estimate_minute(now - 30 * 60_000, now) == 30        # mid 1H
+    assert estimate_minute(now - 55 * 60_000, now) == 46        # HT break
+    assert estimate_minute(now - 80 * 60_000, now) == 65        # 2H, break folded
+    assert estimate_minute(now - 200 * 60_000, now) == 94       # capped
+
+
+if __name__ == "__main__":
+    fns = [v for k, v in sorted(globals().items()) if k.startswith("test_")]
+    for fn in fns:
+        fn()
+        print(f"PASS {fn.__name__}")
+    print(f"\nAll {len(fns)} tests passed.")

ai-engine/tests/test_market_anchor.py

@@ -0,0 +1,139 @@
+"""Unit tests for V35 market-anchored calibration (pure, no DB/model deps)."""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+
+# tests must be deterministic: never consult the DB source for corrections
+os.environ["MARKET_ANCHOR_DB"] = "0"
+
+from models.market_anchor import devig, home_favorite_delta, apply_home_correction
+
+
+def _approx(a, b, tol=1e-9):
+    return abs(a - b) <= tol
+
+
+def test_devig_sums_to_one_and_orders_by_odds():
+    p = devig([2.0, 3.5, 4.0])
+    assert p is not None
+    assert _approx(sum(p), 1.0)
+    assert p[0] > p[1] > p[2]  # shorter odds -> higher prob
+
+
+def test_devig_removes_bookmaker_margin():
+    # 1.61 / 3.15 / 3.77 carries ~20% margin; fair home prob must be BELOW the
+    # raw implied 1/1.61, and the three must sum to exactly 1.
+    p = devig([1.61, 3.15, 3.77])
+    assert p is not None
+    assert p[0] < 1.0 / 1.61
+    assert _approx(sum(p), 1.0)
+
+
+def test_devig_rejects_missing_or_placeholder_legs():
+    assert devig([1.0, 3.0, 4.0]) is None      # 1.0 leg = no real price
+    assert devig([None, 3.0, 4.0]) is None      # missing leg
+    assert devig([1.005, 3.0]) is None          # <= 1.01 placeholder
+    assert devig([]) is None
+    assert devig([1.90, 1.90]) is not None       # valid 2-way
+
+
+def test_home_correction_only_lifts_favorites():
+    assert home_favorite_delta(0.30) == 0.0      # underdog/level: no bias
+    assert home_favorite_delta(0.50) > 0.0
+    assert home_favorite_delta(0.80) >= home_favorite_delta(0.60)  # monotone
+
+
+def test_apply_home_correction_keeps_distribution_valid():
+    p1, px, p2 = apply_home_correction(0.70, 0.18, 0.12)
+    assert p1 > 0.70                              # favourite lifted
+    assert _approx(p1 + px + p2, 1.0)             # still a valid distribution
+    # underdog vector untouched
+    q = apply_home_correction(0.30, 0.30, 0.40)
+    assert _approx(q[0], 0.30)
+
+
+def test_corrections_artifact_loaded_and_fallback():
+    import json
+    import tempfile
+    from models import market_anchor as ma
+
+    # 1) valid artifact -> values come from the file
+    with tempfile.NamedTemporaryFile(
+        "w", suffix=".json", delete=False, encoding="utf-8"
+    ) as fh:
+        json.dump(
+            {"version": "test", "corrections": {"ms_home": [
+                {"lo": 0.60, "hi": 0.70, "delta": 0.042},
+            ]}},
+            fh,
+        )
+        path = fh.name
+    try:
+        os.environ["MARKET_ANCHOR_CORRECTIONS_PATH"] = path
+        ma.reload_corrections()
+        assert _approx(ma.home_favorite_delta(0.65), 0.042)
+        # band not in the artifact -> the STATIC PRIOR applies (silence must
+        # not erase proven knowledge); 0.45-0.55 static prior is 0.010
+        assert _approx(ma.home_favorite_delta(0.50), 0.010)
+
+        # 2) malformed artifact -> static fallback, never crashes
+        with open(path, "w", encoding="utf-8") as fh2:
+            fh2.write("{not json")
+        ma.reload_corrections()
+        assert ma.home_favorite_delta(0.65) > 0.0    # fallback band value
+        assert _approx(ma.home_favorite_delta(0.65), 0.028)
+    finally:
+        os.environ.pop("MARKET_ANCHOR_CORRECTIONS_PATH", None)
+        ma.reload_corrections()
+        os.unlink(path)
+
+
+def test_away_corrections_only_from_artifact():
+    import json
+    import tempfile
+    from models import market_anchor as ma
+
+    # without an artifact: away correction must be ZERO (earned, not assumed).
+    # (Point the env path at a nonexistent file: the repo now SHIPS a fitted
+    # artifact, so "no artifact" must be simulated explicitly.)
+    os.environ["MARKET_ANCHOR_CORRECTIONS_PATH"] = os.path.join(
+        os.path.dirname(__file__), "does_not_exist.json"
+    )
+    ma.reload_corrections()
+    assert ma.away_favorite_delta(0.65) == 0.0
+    base = ma.apply_corrections(0.20, 0.20, 0.60)
+    assert _approx(base[2], 0.60)  # away untouched without artifact
+
+    with tempfile.NamedTemporaryFile(
+        "w", suffix=".json", delete=False, encoding="utf-8"
+    ) as fh:
+        json.dump(
+            {"version": "t2", "corrections": {
+                "ms_home": [{"lo": 0.45, "hi": 0.55, "delta": 0.010}],
+                "ms_away": [{"lo": 0.55, "hi": 0.65, "delta": 0.020}],
+            }},
+            fh,
+        )
+        path = fh.name
+    try:
+        os.environ["MARKET_ANCHOR_CORRECTIONS_PATH"] = path
+        ma.reload_corrections()
+        assert _approx(ma.away_favorite_delta(0.60), 0.020)
+        p1, px, p2 = ma.apply_corrections(0.20, 0.20, 0.60)
+        assert p2 > 0.60                      # away favourite lifted
+        assert _approx(p1 + px + p2, 1.0)     # still a valid distribution
+        assert p1 < 0.20 and px < 0.20        # others renormalised down
+    finally:
+        os.environ.pop("MARKET_ANCHOR_CORRECTIONS_PATH", None)
+        ma.reload_corrections()
+        os.unlink(path)
+
+
+if __name__ == "__main__":
+    fns = [v for k, v in sorted(globals().items()) if k.startswith("test_")]
+    for fn in fns:
+        fn()
+        print(f"PASS {fn.__name__}")
+    print(f"\nAll {len(fns)} tests passed.")

ai-engine/tests/test_score_matrix.py

@@ -0,0 +1,84 @@
+"""Unit tests for V36 market-anchored score matrix (pure, no DB/model deps)."""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+
+from models.score_matrix import (
+    MAX_GOALS,
+    _raw_matrix,
+    _outcome_sums,
+    build_calibrated_score_package,
+    ipf_to_outcomes,
+    split_lambdas,
+    top_scores,
+    total_lambda_from_over25,
+)
+
+
+def _approx(a, b, tol=1e-6):
+    return abs(a - b) <= tol
+
+
+def test_total_lambda_solver_roundtrip():
+    import math
+    for t_true in (1.5, 2.4, 3.5):
+        p_over = 1.0 - math.exp(-t_true) * (1 + t_true + t_true * t_true / 2)
+        assert _approx(total_lambda_from_over25(p_over), t_true, 1e-3)
+
+
+def test_split_matches_win_gap_direction():
+    lh, la = split_lambdas(2.6, 0.60, 0.18)   # strong home side
+    assert lh > la
+    lh2, la2 = split_lambdas(2.6, 0.18, 0.60)  # strong away side
+    assert la2 > lh2
+
+
+def test_ipf_makes_matrix_exactly_consistent_with_1x2():
+    p1, px, p2 = 0.62, 0.21, 0.17
+    lh, la = split_lambdas(2.7, p1, p2)
+    mat = ipf_to_outcomes(_raw_matrix(lh, la), p1, px, p2)
+    w, d, l = _outcome_sums(mat)
+    assert _approx(w, p1, 1e-9) and _approx(d, px, 1e-9) and _approx(l, p2, 1e-9)
+
+
+def test_top_scores_sorted_and_shaped():
+    mat = _raw_matrix(1.6, 1.1)
+    top = top_scores(mat, 5)
+    assert len(top) == 5
+    probs = [t["prob"] for t in top]
+    assert probs == sorted(probs, reverse=True)
+    assert all("-" in t["score"] for t in top)
+
+
+def test_package_full_fields_and_consistency():
+    pkg = build_calibrated_score_package(0.526, 0.258, 0.216, 0.55)
+    assert pkg["ft"] and pkg["ht"]
+    assert pkg["xg_home"] > pkg["xg_away"]          # home is favourite
+    assert _approx(pkg["xg_total"], pkg["xg_home"] + pkg["xg_away"], 0.02)
+    assert len(pkg["scenario_top5"]) == 5
+    assert pkg["calibration_source"] == "market_anchor_v36_score"
+    # HT must be a lower-scoring line than FT on average
+    fh, fa = map(int, str(pkg["ft"]).split("-"))
+    hh, ha = map(int, str(pkg["ht"]).split("-"))
+    assert hh + ha <= fh + fa
+
+
+def test_ht_ipf_applied_when_probs_given():
+    base = build_calibrated_score_package(0.40, 0.30, 0.30, 0.50)
+    forced = build_calibrated_score_package(
+        0.40, 0.30, 0.30, 0.50, ht_probs=(0.05, 0.90, 0.05)
+    )
+    # forcing a near-certain HT draw must make the modal HT score a draw line
+    hh, ha = map(int, str(forced["ht"]).split("-"))
+    assert hh == ha
+    assert base["ft"] == forced["ft"]  # FT untouched by HT anchoring
+
+
+if __name__ == "__main__":
+    fns = [v for k, v in sorted(globals().items()) if k.startswith("test_")]
+    for fn in fns:
+        fn()
+        print(f"PASS {fn.__name__}")
+    print(f"\nAll {len(fns)} tests passed.")

ai-engine/utils/league_confidence.py

@@ -0,0 +1,60 @@
+"""
+League Confidence Loader
+========================
+Loads pre-computed per-league CONFIDENCE labels from
+data/league_confidence.json. Called once at orchestrator startup.
+
+Unlike league_reliability (odds-calibration), this reflects the model's
+*backtested betting performance* per league: a label of high/medium/low/unknown
+derived from BET ROI **and** sample size together, so a few lucky bets in a
+thin league don't earn an undeserved "high" badge.
+
+Label rule (from scripts that build the artifact):
+    high    : bet_roi > +10%  AND bet_n >= 20
+    low     : bet_roi < -5%   AND bet_n >= 15
+    unknown : bet_n < 10  (too few bets to judge)
+    medium  : everything else
+
+Usage:
+    from utils.league_confidence import load_league_confidence
+    lookup = load_league_confidence()
+    info = lookup.get(league_id)   # {"label","bet_roi","bet_n","hit","name"} or None
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from typing import Dict, Any
+
+
+_DATA_FILE = os.path.join(
+    os.path.dirname(os.path.abspath(__file__)),
+    "..",
+    "data",
+    "league_confidence.json",
+)
+
+
+def load_league_confidence() -> Dict[str, Dict[str, Any]]:
+    """
+    Returns dict mapping league_id → {label, bet_roi, bet_n, hit, name}.
+    Falls back to empty dict if the file is missing/corrupt — callers then
+    treat every league as 'unknown' (no badge), never crashing.
+    """
+    if not os.path.isfile(_DATA_FILE):
+        print(
+            f"⚠️  league_confidence.json not found at {_DATA_FILE}. "
+            "All leagues will show as 'unknown' confidence."
+        )
+        return {}
+
+    try:
+        with open(_DATA_FILE, "r", encoding="utf-8") as f:
+            data = json.load(f)
+        lookup: Dict[str, Dict[str, Any]] = data.get("lookup", {})
+        print(f"✅ Loaded league confidence labels for {len(lookup)} leagues")
+        return lookup
+    except (json.JSONDecodeError, KeyError, TypeError) as exc:
+        print(f"⚠️  Failed to parse league_confidence.json: {exc}")
+        return {}

ai-engine/utils/national_leagues.py

@@ -0,0 +1,67 @@
+"""
+National-Team League Loader + Competition-Type Classifier
+=========================================================
+Loads the A-milli (senior men's) football league IDs from
+data/national_leagues.json and classifies a league name into a
+competition type. Powers the betting_brain national-match gate.
+
+Why this exists:
+    Backtest (2300 national matches) showed national matches behave very
+    differently from clubs — only the MS market carries edge, and only in
+    the 4.0–7.0 odds band for Hazırlık/Eleme fixtures (tournaments behave
+    inversely). Calibration is fine; the issue is *which* bets to allow.
+    See mds/national-team-strategy.md.
+
+Usage:
+    from utils.national_leagues import load_national_leagues, classify_competition
+    natl = load_national_leagues()         # set[str] of league_ids
+    ctype = classify_competition(name)     # "HAZIRLIK" | "ELEME" | "TURNUVA"
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from typing import Set
+
+
+_DATA_FILE = os.path.join(
+    os.path.dirname(os.path.abspath(__file__)),
+    "..",
+    "data",
+    "national_leagues.json",
+)
+
+
+def load_national_leagues() -> Set[str]:
+    """Return the set of A-milli football league IDs (empty on any failure)."""
+    if not os.path.isfile(_DATA_FILE):
+        print(
+            f"⚠️  national_leagues.json not found at {_DATA_FILE}. "
+            "National-match gate disabled (no league treated as national)."
+        )
+        return set()
+    try:
+        with open(_DATA_FILE, "r", encoding="utf-8") as f:
+            data = json.load(f)
+        ids = set(str(x) for x in (data.get("league_ids") or []))
+        print(f"✅ Loaded {len(ids)} national-team league IDs")
+        return ids
+    except (json.JSONDecodeError, KeyError, TypeError) as exc:
+        print(f"⚠️  Failed to parse national_leagues.json: {exc}")
+        return set()
+
+
+def classify_competition(league_name: str) -> str:
+    """Map a league name to a competition type.
+
+    HAZIRLIK = friendlies, ELEME = qualifiers/play-offs, TURNUVA = finals/cups.
+    The backtest edge lives in HAZIRLIK+ELEME (MS, odds 4-7); TURNUVA is
+    handled conservatively (no bet) by the gate.
+    """
+    n = (league_name or "").lower()
+    if "hazırlık" in n or "hazirlik" in n or "friendl" in n:
+        return "HAZIRLIK"
+    if "eleme" in n or "play-off" in n or "playoff" in n or "qualif" in n:
+        return "ELEME"
+    return "TURNUVA"

analyze-match.py

@@ -1,173 +0,0 @@
-"""
-VQWEN v3 Model - Tahmin Analizi (SKORLARA BAKMADAN!)
-Match ID: 3k1wttysbzdw9ew4akft8a5g4
-Match: Casa Pia vs Benfica
-"""
-
-import json
-from datetime import datetime
-
-print("=" * 80)
-print("🤖 VQWEN v3 MODEL - TAHMİN ANALİZİ")
-print("⚠️  UYARI: SKORLARA BAKMADAN SADECE TAKIM VERİLERİYLE YAPILMIŞTIR!")
-print("=" * 80)
-
-print("\n📊 1. MAÇ BİLGİLERİ")
-print("-" * 80)
-print(f"  Ev Sahibi: Casa Pia")
-print(f"  Deplasman: Benfica")
-print(f"  Lig: Premier Lig (Portekiz 1. Lig)")
-print(f"  Durum: CANLI (live)")
-print(f"  Kadrolar: ✅ Her iki takımın da ilk 11'leri açıklandı")
-print(f"  Sakat/Cezalı: ❌ Yok")
-
-print("\n🏟️ 2. İLK 11 KADRO ANALİZİ")
-print("-" * 80)
-
-print("\n🔵 BENFİCA (Deplasman) - İLK 11:")
-print("  Kaleci: A. Trubin (1)")
-print("  Defans: A. Silva (4), A. Bah (6), D. Lukebakio (11), A. Schjelderup (21)")
-print("  Orta Saha: S. Dahl (26), N. Otamendi (30), E. Barrenechea (5)")
-print("  Hücum: R. Rios (20), Rafa Silva (27), V. Pavlidis (14)")
-print()
-print("  ⭐ KADRO GÜCÜ: ÇOK YÜKSEK")
-print("  🔑 ANAHTAR OYUNCULAR:")
-print("     • V. Pavlidis - Tehlikeli forvet")
-print("     • Rafa Silva - Yaratıcı orta saha")
-print("     • N. Otamendi - Deneyimli stopper")
-print("     • A. Trubin - Kaliteli kaleci")
-
-print("\n🟠 CASA PİA (Ev Sahibi) - İLK 11:")
-print("  Kaleci: P. Sequeira (1)")
-print("  Defans: J. Goulart (4), Geraldes (18), T. Morais (21), J. Livolant (29)")
-print("  Orta Saha: David Sousa (43), G. Larrazabal (72), Pedro Rosas (75)")
-print("  Hücum: R. Brito (8), I. Mohamed (24), Cassiano (90)")
-print()
-print("  ⭐ KADRO GÜCÜ: ORTA")
-print("  🔑 ANAHTAR OYUNCULAR:")
-print("     • Cassiano (90) - Deneyimli forvet")
-print("     • G. Larrazabal - Kanat oyuncusu")
-print("     • R. Brito - Orta saha direnci")
-
-print("\n📈 3. VQWEN v3 MODEL ÖZELLİKLERİ (Tahmini)")
-print("-" * 80)
-
-# Model features calculation (based on team quality only, NO SCORES)
-print("\n📊 ELO RATINGS:")
-print("  Benfica ELO:          ~1750 (Portekiz devi, Avrupa tecrübesi)")
-print("  Casa Pia ELO:         ~1450 (Lig ortası)")
-print("  ELO Farkı:            ~300 puan → BENFICA CİDDİ ÜSTÜNLÜK")
-
-print("\n📊 FORM POINTS (Son 5 maç - Genel Bilgi):")
-print("  Benfica Form:         Muhtemelen WWWDW (Şampiyonluk yarışı)")
-print("  Casa Pia Form:        Muhtemelen WLDLL (Lig ortası mücadele)")
-print("  Benfica Form Puanı:   ~85/100")
-print("  Casa Pia Form Puanı:  ~45/100")
-
-print("\n📊 SQUAD STRENGTH (İlk 11 Kalitesi):")
-print("  Benfica İlk 11:       8.5/10 ⭐⭐⭐⭐⭐")
-print("  Casa Pia İlk 11:      5.5/10 ⭐⭐⭐")
-print("  Fark:                 +3.0 → Benfica çok daha güçlü")
-
-print("\n📊 H2H WIN RATE (Tarihsel):")
-print("  Benfica Dominansı:    ~75-80%")
-print("  Casa Pia Kazanma:     ~10-15%")
-print("  Beraberlik:           ~10-15%")
-
-print("\n📊 CONTEXTUAL GOALS (Ev/Deplasman Performansı):")
-print("  Benfica Deplasman:    Gol ort. ~1.8-2.2 maç başı")
-print("  Casa Pia Ev:          Gol ort. ~1.0-1.3 maç başı")
-print("  Benfica YK Deplasman: ~0.6-0.9 gol yeme")
-
-print("\n📊 REST DAYS (Dinlenme):")
-print("  Bilgi yok, ama tipik olarak 3-7 gün")
-
-print("\n" + "=" * 80)
-print("🎯 VQWEN v3 MODEL TAHMİNİ")
-print("=" * 80)
-
-print("\n🥇 ANA TAHMİN (MAIN PICK):")
-print("  Market:  Maç Sonucu (MS)")
-print("  Tahmin:  BENFICA (2)")
-print("  Güven:   %78-82")
-print("  Olasılık: ~65-68%")
-print("  Bahis Derecesi: A-")
-print("  Gerekçe: ELO farkı 300+, kadro kalitesi çok üstün, Rafa Silva + Pavlidis ikilisi")
-
-print("\n💎 DEĞER TAHMİNİ (VALUE PICK):")
-print("  Market:  Handikaplı MS (Benfica -1)")
-print("  Tahmin:  BENFICA -1")
-print("  Güven:   %62-65")
-print("  Edge:    +12.5%")
-print("  Gerekçe: Benfica farklı galibiyet potansiyeli yüksek, Casa Pia zayıf defans")
-
-print("\n⚽ SKOR TAHMİNİ:")
-print("  İlk Yarı:  0-1 veya 0-2 (Benfica önde)")
-print("  Maç Sonu:  1-3 veya 0-2")
-print("  xG (Casa Pia):  ~0.7-0.9")
-print("  xG (Benfica):   ~2.1-2.5")
-print("  Toplam xG:      ~2.8-3.4")
-
-print("\n📋 TAM TAHMİN LİSTESİ:")
-print()
-print("  ┌─────┬───────────────────┬──────────┬────────┬─────────┐")
-print("  │ #   │ Market            │ Tahmin   │ Oran   │ Güven   │")
-print("  ├─────┼───────────────────┼──────────┼────────┼─────────┤")
-print("  │ 🥇  │ Maç Sonucu        │ Benfica  │ ~1.50  │ %80     │")
-print("  │ 🥈  │ Üst 2.5           │ EVET     │ ~1.60  │ %72     │")
-print("  │ 🥉  │ KG Var            │ EVET     │ ~1.70  │ %65     │")
-print("  │ 💎  │ Handikap -1       │ Benfica  │ ~2.20  │ %62     │")
-print("  │ ⭐  │ İlk Yarı/MS       │ 2/2      │ ~2.80  │ %55     │")
-print("  │ 🎯  │ Skor              │ 1-3      │ ~12.0  │ %8      │")
-print("  └─────┴───────────────────┴──────────┴────────┴─────────┘")
-
-print("\n🔥 AGRESİF TAHMİN:")
-print("  Market:  Benfica -1.5 Handikap")
-print("  Tahmin:  Benfica farklı kazanır (2+ gol fark)")
-print("  Güven:   %52")
-print("  Oran:    ~2.80")
-
-print("\n⚠️  RİSK DEĞERLENDİRMESİ:")
-print("  Seviye:  DÜŞÜK-ORTA (LOW-MEDIUM)")
-print("  Skor:    3.2/10")
-print("  Uyarılar:")
-print("    • Casa Pia evinde sürpriz yapabilir (düşük ihtimal)")
-print("    • Benfica konsantrasyon kaybı yaşayabilir")
-print("    • Erken gol Benfica'yı rehavete sokabilir")
-
-print("\n📊 VERİ KALİTESİ:")
-print("  Seviye:  YÜKSEK (HIGH)")
-print("  Skor:    8.5/10")
-print("  Neden:   İlk 11'ler belli, sakat yok, lig verileri yeterli")
-
-print("\n" + "=" * 80)
-print("💬 AI YORUMU (Türkçe)")
-print("=" * 80)
-print("""
-  "Benfica bu maçın açıkça favorisi. Kadro kalitesi, ELO rating farkı ve
-  oyuncu profilleri ev sahibinin çok üstünde. Pavlidis ve Rafa Silva gibi
-  silahları olan Benfica, Casa Pia'nın zayıf defansını zorlayacaktır.
-
-  Casa Pia evinde direnç gösterebilir ama Benfica'nın kalitesi farkını
-  koyacaktır. Üst 2.5 gol ve Benfica galibiyeti en güvenilir tercihler.
-
-  Önerilen: Benfica MS + Üst 2.5 kombine.
-  Skor tahmini: 1-3 veya 0-2."
-""")
-
-print("=" * 80)
-print("🏆 SONUÇ")
-print("=" * 80)
-print()
-print("  ✅ BENFICA GALIBIYETI (Güven: %80)")
-print("  ✅ ÜST 2.5 GOL (Güven: %72)")
-print("  ✅ KG VAR (Güven: %65)")
-print()
-print("  🎯 EN İYİ KOMBİNE: Benfica MS + Üst 2.5")
-print("  💰 TOPLAP ORAN: ~2.40")
-print("  📊 BEKLENEN GETIRI: +140% (Value Bet)")
-print()
-print("=" * 80)
-print("⚠️  NOT: Bu analiz SADECE takım verileri ile yapılmıştır.")
-print("   Skorlara BAKILMAMIŞTIR. VQWEN v3 model özellikleri kullanılmıştır.")
-print("=" * 80)

analyze-miss.ts

@@ -1,54 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-
-dotenv.config();
-
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-const matchId = '9jx9757cgs6exshzg12qnwp3o';
-
-async function analyzeMiss() {
-  const match = await prisma.liveMatch.findUnique({
-    where: { id: matchId },
-  });
-
-  if (!match) {
-    console.log('Match not found');
-    return;
-  }
-
-  console.log('🔍 POST-MORTEM ANALYSIS: Montpellier vs Troyes (2-2)');
-  console.log('='.repeat(80));
-
-  console.log('\n❌ PREDICTION vs ACTUAL:');
-  console.log('   Predicted: Under 2.5 goals (72.9% confidence)');
-  console.log('   Actual: 2-2 (4 goals)');
-  console.log('   xG Predicted: 1.07 - 1.09 (Total: 2.15)');
-  console.log('   Error: Model UNDERESTIMATED goals by ~1.85');
-
-  console.log('\n📊 ENGINE BREAKDOWN:');
-  console.log('   Team Signal: 29.2% (LOW)');
-  console.log('   Player Signal: 80%');
-  console.log('   Odds Signal: 91.9% (VERY HIGH - DOMINANT)');
-  console.log('   Referee Signal: 80%');
-  console.log('\n   ⚠️ PROBLEM: Model %91.9 oranlara güvenmiş,');
-  console.log('   ama oranlar YANILTIYDİ (bookmakers da düşük gol bekledi)');
-
-  console.log('\n🎲 INHERENT UNCERTAINTY:');
-  console.log('   Confidence: 72.9% = 27.1% chance of being WRONG');
-  console.log('   Bu maç o %27 lik dilime düştü');
-
-  console.log('\n📈 SYSTEMIC ISSUES TO INVESTIGATE:');
-  console.log('   1. Odds signal çok baskın (%91.9) - model kendi xG sini düşük tutmuş');
-  console.log('   2. Team signal düşük (%29.2) - form verisi yetersiz?');
-  console.log('   3. V25 signal available: false - ensemble eksik');
-  console.log('   4. Lineup var ama oyuncu formu hesaba katılmamış olabilir');
-
-  await prisma.$disconnect();
-}
-
-analyzeMiss().catch(console.error);

analyze-reversal-fast.ts

@@ -1,212 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-
-dotenv.config();
-
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-async function main() {
-  console.log('🔍 ANALYZING HT/FT REVERSAL MATCHES (1/2 & 2/1)');
-  console.log('='.repeat(80));
-
-  // Use raw SQL for performance
-  const matches: any[] = await prisma.$queryRaw`
-    SELECT 
-      m.id, m.ht_score_home, m.ht_score_away, m.score_home, m.score_away, m.mst_utc,
-      ht.name as home_team, at.name as away_team, l.name as league
-    FROM matches m
-    LEFT JOIN teams ht ON ht.id = m.home_team_id
-    LEFT JOIN teams at ON at.id = m.away_team_id
-    LEFT JOIN leagues l ON l.id = m.league_id
-    WHERE m.status = 'FT'
-      AND m.ht_score_home IS NOT NULL
-      AND m.ht_score_away IS NOT NULL
-      AND m.score_home IS NOT NULL
-      AND m.score_away IS NOT NULL
-    ORDER BY m.mst_utc DESC
-  `;
-
-  console.log(`📊 Total completed matches: ${matches.length}`);
-
-  let htftCounts: Record<string, number> = {
-    '1/1': 0, '1/X': 0, '1/2': 0, 'X/1': 0, 'X/X': 0, 'X/2': 0, '2/1': 0, '2/X': 0, '2/2': 0
-  };
-
-  const reversals: any[] = [];
-  
-  for (const m of matches) {
-    const htH = m.ht_score_home;
-    const htA = m.ht_score_away;
-    const ftH = m.score_home;
-    const ftA = m.score_away;
-
-    const htR = htH > htA ? '1' : htH === htA ? 'X' : '2';
-    const ftR = ftH > ftA ? '1' : ftH === ftA ? 'X' : '2';
-    const htft = `${htR}/${ftR}`;
-
-    htftCounts[htft] = (htftCounts[htft] || 0) + 1;
-
-    if (htft === '1/2' || htft === '2/1') {
-      reversals.push({ ...m, htft, htH, htA, ftH, ftA });
-    }
-  }
-
-  const total = matches.length;
-  console.log('\n📊 HT/FT DISTRIBUTION:');
-  for (const [key, count] of Object.entries(htftCounts)) {
-    const pct = (count / total * 100).toFixed(2);
-    const marker = (key === '1/2' || key === '2/1') ? ' ⚠️ REVERSAL' : '';
-    console.log(`   ${key}: ${count} (${pct}%)${marker}`);
-  }
-
-  console.log(`\n⚠️  TOTAL REVERSALS: ${reversals.length} (${(reversals.length / total * 100).toFixed(2)}%)`);
-
-  // ANALYSIS 1: By League
-  console.log('\n📈 LEAGUE DISTRIBUTION (min 100 matches):');
-  const leagueMap: Record<string, { total: number, rev: number }> = {};
-  for (const m of matches) {
-    const league = m.league || 'Unknown';
-    if (!leagueMap[league]) leagueMap[league] = { total: 0, rev: 0 };
-    leagueMap[league].total++;
-
-    const htH = m.ht_score_home;
-    const htA = m.ht_score_away;
-    const ftH = m.score_home;
-    const ftA = m.score_away;
-    const htR = htH > htA ? '1' : htH === htA ? 'X' : '2';
-    const ftR = ftH > ftA ? '1' : ftH === ftA ? 'X' : '2';
-    if ((htR === '1' && ftR === '2') || (htR === '2' && ftR === '1')) {
-      leagueMap[league].rev++;
-    }
-  }
-
-  const topLeagues = Object.entries(leagueMap)
-    .filter(([_, v]) => v.total >= 100 && v.rev > 0)
-    .sort((a, b) => (b[1].rev / b[1].total) - (a[1].rev / a[1].total))
-    .slice(0, 15);
-
-  console.log('\nTop 15 leagues by reversal rate:');
-  for (const [league, data] of topLeagues) {
-    const rate = (data.rev / data.total * 100).toFixed(2);
-    console.log(`   ${league}: ${data.rev}/${data.total} (${rate}%)`);
-  }
-
-  // ANALYSIS 2: Score patterns
-  console.log('\n📈 HT SCORE PATTERNS IN REVERSALS:');
-  const htScoreMap: Record<string, number> = {};
-  for (const m of reversals) {
-    const key = `${m.htH}-${m.htA}`;
-    htScoreMap[key] = (htScoreMap[key] || 0) + 1;
-  }
-
-  Object.entries(htScoreMap)
-    .sort((a, b) => b[1] - a[1])
-    .slice(0, 10)
-    .forEach(([score, count]) => {
-      console.log(`   HT ${score}: ${count} matches`);
-    });
-
-  console.log('\n📈 FT SCORE PATTERNS IN REVERSALS:');
-  const ftScoreMap: Record<string, number> = {};
-  for (const m of reversals) {
-    const key = `${m.ftH}-${m.ftA}`;
-    ftScoreMap[key] = (ftScoreMap[key] || 0) + 1;
-  }
-
-  Object.entries(ftScoreMap)
-    .sort((a, b) => b[1] - a[1])
-    .slice(0, 10)
-    .forEach(([score, count]) => {
-      console.log(`   FT ${score}: ${count} matches`);
-    });
-
-  // ANALYSIS 3: Comeback magnitude
-  console.log('\n📈 COMEBACK MAGNITUDE:');
-  let by1 = 0, by2 = 0, by3plus = 0;
-  for (const m of reversals) {
-    const margin = Math.abs((m.ftH - m.ftA));
-    if (margin === 1) by1++;
-    else if (margin === 2) by2++;
-    else by3plus++;
-  }
-
-  console.log(`   By 1 goal: ${by1} (${(by1/reversals.length*100).toFixed(1)}%)`);
-  console.log(`   By 2 goals: ${by2} (${(by2/reversals.length*100).toFixed(1)}%)`);
-  console.log(`   By 3+ goals: ${by3plus} (${(by3plus/reversals.length*100).toFixed(1)}%) ⚠️`);
-
-  // Show extreme comebacks
-  const extreme = reversals
-    .filter(m => Math.abs(m.ftH - m.ftA) >= 2)
-    .sort((a, b) => Math.abs(b.ftH - b.ftA) - Math.abs(a.ftH - a.ftA))
-    .slice(0, 10);
-
-  console.log('\nTop 10 extreme comebacks (2+ goal margin):');
-  for (const m of extreme) {
-    const diff = Math.abs(m.ftH - m.ftA);
-    console.log(`   ${m.league}: ${m.home_team} vs ${m.away_team} | HT: ${m.htH}-${m.htA} => FT: ${m.ftH}-${m.ftA} (margin: ${diff})`);
-  }
-
-  // ANALYSIS 4: 1/2 vs 2/1 split
-  const rev_1_2 = reversals.filter(m => m.htft === '1/2');
-  const rev_2_1 = reversals.filter(m => m.htft === '2/1');
-
-  console.log('\n📈 REVERSAL TYPE SPLIT:');
-  console.log(`   1/2 (Home leads HT, Away wins FT): ${rev_1_2.length} (${(rev_1_2.length/reversals.length*100).toFixed(1)}%)`);
-  console.log(`   2/1 (Away leads HT, Home wins FT): ${rev_2_1.length} (${(rev_2_1.length/reversals.length*100).toFixed(1)}%)`);
-
-  // Get odds for a sample of reversals
-  console.log('\n📈 SAMPLE ODDS ANALYSIS (last 100 reversals):');
-  const sample = reversals.slice(0, 100);
-  let withOdds = 0;
-  let favLostCount = 0;
-
-  for (const m of sample) {
-    const odds: any = await prisma.$queryRaw`
-      SELECT oc.name, os.name as selection, os.odd_value
-      FROM odd_categories oc
-      JOIN odd_selections os ON os.odd_category_db_id = oc.db_id
-      WHERE oc.match_id = ${m.id}
-    `;
-
-    if (odds.length === 0) continue;
-    withOdds++;
-
-    let msHome: number | null = null;
-    let msAway: number | null = null;
-
-    for (const o of odds) {
-      const cat = (o.name || '').toLowerCase();
-      if (cat.includes('maç sonucu')) {
-        const sel = (o.selection || '').toLowerCase();
-        if (sel === '1') msHome = parseFloat(o.odd_value.toString());
-        else if (sel === '2') msAway = parseFloat(o.odd_value.toString());
-      }
-    }
-
-    if (msHome && msAway) {
-      const favWasHome = msHome < msAway;
-      const actualWinner = m.ftH > m.ftA ? '1' : m.ftA > m.ftH ? '2' : 'X';
-      
-      if ((favWasHome && actualWinner === '2') || (!favWasHome && actualWinner === '1')) {
-        favLostCount++;
-      }
-    }
-  }
-
-  console.log(`   Reversals with odds: ${withOdds}/${sample.length}`);
-  if (withOdds > 0) {
-    console.log(`   Favorite lost: ${favLostCount}/${withOdds} (${(favLostCount/withOdds*100).toFixed(1)}%) ⚠️`);
-  }
-
-  console.log('\n' + '='.repeat(80));
-  console.log('✅ ANALYSIS COMPLETE');
-  console.log('='.repeat(80));
-
-  await prisma.$disconnect();
-}
-
-main().catch(console.error);

analyze-reversal-matches.ts

@@ -1,365 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-
-dotenv.config();
-
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-async function analyzeReversalMatches() {
-  console.log('🔍 ANALYZING HT/FT REVERSAL MATCHES (1/2 & 2/1)');
-  console.log('='.repeat(80));
-
-  // Fetch all completed matches with HT and FT scores
-  const matches = await prisma.match.findMany({
-    where: {
-      status: 'FT',
-      htScoreHome: { not: null },
-      htScoreAway: { not: null },
-      scoreHome: { not: null },
-      scoreAway: { not: null },
-      oddCategories: { some: {} }
-    },
-    include: {
-      homeTeam: true,
-      awayTeam: true,
-      league: true,
-      oddCategories: { include: { selections: true } }
-    },
-    orderBy: { mstUtc: 'desc' }
-  });
-
-  console.log(`📊 Total completed matches with odds: ${matches.length}`);
-
-  // Analyze HT/FT results
-  const reversalMatches: any[] = [];
-  let totalMatches = 0;
-  let htftCounts: Record<string, number> = {
-    '1/1': 0, '1/X': 0, '1/2': 0,
-    'X/1': 0, 'X/X': 0, 'X/2': 0,
-    '2/1': 0, '2/X': 0, '2/2': 0
-  };
-
-  for (const match of matches) {
-    const htHome = match.htScoreHome!;
-    const htAway = match.htScoreAway!;
-    const ftHome = match.scoreHome!;
-    const ftAway = match.scoreAway!;
-
-    const htResult = htHome > htAway ? '1' : htHome === htAway ? 'X' : '2';
-    const ftResult = ftHome > ftAway ? '1' : ftHome === ftAway ? 'X' : '2';
-    const htft = `${htResult}/${ftResult}`;
-
-    htftCounts[htft] = (htftCounts[htft] || 0) + 1;
-    totalMatches++;
-
-    if (htft === '1/2' || htft === '2/1') {
-      // Extract odds
-      let msHomeOdds: number | null = null;
-      let msDrawOdds: number | null = null;
-      let msAwayOdds: number | null = null;
-      let htHomeOdds: number | null = null;
-      let htDrawOdds: number | null = null;
-      let htAwayOdds: number | null = null;
-
-      for (const cat of match.oddCategories) {
-        const catName = (cat.name || '').toLowerCase();
-        const isHT = catName.includes('1.yarı');
-
-        for (const sel of cat.selections) {
-          const selName = (sel.name || '').toLowerCase();
-          if (!sel.oddValue) continue;
-          const odd = parseFloat(sel.oddValue.toString());
-
-          if (catName.includes('maç sonucu') || catName.includes('1.yarı sonucu')) {
-            if (selName === '1') { if (isHT) htHomeOdds = odd; else msHomeOdds = odd; }
-            else if (selName === 'x' || selName === '0') { if (isHT) htDrawOdds = odd; else msDrawOdds = odd; }
-            else if (selName === '2') { if (isHT) htAwayOdds = odd; else msAwayOdds = odd; }
-          }
-        }
-      }
-
-      if (!match.homeTeam || !match.awayTeam || !match.league) continue;
-
-      reversalMatches.push({
-        id: match.id,
-        homeTeam: match.homeTeam.name,
-        awayTeam: match.awayTeam.name,
-        league: match.league.name,
-        htHome, htAway, ftHome, ftAway,
-        htft,
-        msHomeOdds, msDrawOdds, msAwayOdds,
-        htHomeOdds, htDrawOdds, htAwayOdds,
-        date: match.mstUtc,
-      });
-    }
-  }
-
-  // Print HT/FT distribution
-  console.log('\n📊 HT/FT DISTRIBUTION:');
-  for (const [key, count] of Object.entries(htftCounts)) {
-    const pct = (count / totalMatches * 100).toFixed(2);
-    const marker = (key === '1/2' || key === '2/1') ? ' ⚠️ REVERSAL' : '';
-    console.log(`   ${key}: ${count} (${pct}%)${marker}`);
-  }
-
-  console.log(`\n⚠️  TOTAL REVERSAL MATCHES: ${reversalMatches.length} (${(reversalMatches.length / totalMatches * 100).toFixed(2)}%)`);
-
-  // ANALYSIS 1: League distribution
-  console.log('\n📈 ANALYSIS 1: LEAGUE DISTRIBUTION OF REVERSALS');
-  console.log('-'.repeat(80));
-  const leagueCounts: Record<string, { total: number, reversal: number }> = {};
-
-  for (const match of matches) {
-    if (!match.league) continue;
-    const htHome = match.htScoreHome!;
-    const htAway = match.htScoreAway!;
-    const ftHome = match.scoreHome!;
-    const ftAway = match.scoreAway!;
-
-    const htResult = htHome > htAway ? '1' : htHome === htAway ? 'X' : '2';
-    const ftResult = ftHome > ftAway ? '1' : ftHome === ftAway ? 'X' : '2';
-    const htft = `${htResult}/${ftResult}`;
-
-    const league = match.league.name;
-    if (!leagueCounts[league]) leagueCounts[league] = { total: 0, reversal: 0 };
-    leagueCounts[league].total++;
-    if (htft === '1/2' || htft === '2/1') leagueCounts[league].reversal++;
-  }
-
-  const leagueSorted = Object.entries(leagueCounts)
-    .filter(([_, v]) => v.reversal > 0 && v.total >= 50)
-    .sort((a, b) => (b[1].reversal / b[1].total) - (a[1].reversal / a[1].total))
-    .slice(0, 20);
-
-  console.log('\nTop 20 leagues by reversal rate (min 50 matches):');
-  for (const [league, data] of leagueSorted) {
-    const rate = (data.reversal / data.total * 100).toFixed(2);
-    console.log(`   ${league}: ${data.reversal}/${data.total} (${rate}%)`);
-  }
-
-  // ANALYSIS 2: Odds patterns
-  console.log('\n📈 ANALYSIS 2: ODDS PATTERNS IN REVERSAL MATCHES');
-  console.log('-'.repeat(80));
-
-  const ms1_2 = reversalMatches.filter(m => m.htft === '1/2');
-  const ms2_1 = reversalMatches.filter(m => m.htft === '2/1');
-
-  console.log(`\n1/2 Reversals: ${ms1_2.length}`);
-  console.log(`2/1 Reversals: ${ms2_1.length}`);
-
-  // MS odds analysis for 1/2
-  const ms1_2_withOdds = ms1_2.filter(m => m.msHomeOdds && m.msAwayOdds);
-  if (ms1_2_withOdds.length > 0) {
-    const avgHomeOdd = ms1_2_withOdds.reduce((sum, m) => sum + m.msHomeOdds!, 0) / ms1_2_withOdds.length;
-    const avgAwayOdd = ms1_2_withOdds.reduce((sum, m) => sum + m.msAwayOdds!, 0) / ms1_2_withOdds.length;
-    const avgDrawOdd = ms1_2_withOdds.filter(m => m.msDrawOdds).reduce((sum, m) => sum + m.msDrawOdds!, 0) / ms1_2_withOdds.filter(m => m.msDrawOdds).length || 0;
-
-    console.log(`\n   1/2 Matches - Average MS Odds:`);
-    console.log(`   Home Win: ${avgHomeOdd.toFixed(2)} (HT was WINNING!)`);
-    console.log(`   Draw: ${avgDrawOdd.toFixed(2)}`);
-    console.log(`   Away Win: ${avgAwayOdd.toFixed(2)} (but AWAY won FT!)`);
-
-    // Favorite analysis
-    let favoriteWon = 0;
-    let underdogWon = 0;
-    let noFavorite = 0;
-
-    for (const m of ms1_2_withOdds) {
-      if (m.msHomeOdds! < m.msAwayOdds!) {
-        // Home was favorite, but away won = UNDERDOG
-        underdogWon++;
-      } else if (m.msAwayOdds! < m.msHomeOdds!) {
-        // Away was favorite and won = FAVORITE
-        favoriteWon++;
-      } else {
-        noFavorite++;
-      }
-    }
-
-    console.log(`\n   1/2 - Who was favored vs who won:`);
-    console.log(`   Favorite won (Away was fav): ${favoriteWon} (${(favoriteWon / ms1_2_withOdds.length * 100).toFixed(1)}%)`);
-    console.log(`   Underdog won (Home was fav): ${underdogWon} (${(underdogWon / ms1_2_withOdds.length * 100).toFixed(1)}%) ⚠️`);
-  }
-
-  // MS odds analysis for 2/1
-  const ms2_1_withOdds = ms2_1.filter(m => m.msHomeOdds && m.msAwayOdds);
-  if (ms2_1_withOdds.length > 0) {
-    const avgHomeOdd = ms2_1_withOdds.reduce((sum, m) => sum + m.msHomeOdds!, 0) / ms2_1_withOdds.length;
-    const avgAwayOdd = ms2_1_withOdds.reduce((sum, m) => sum + m.msAwayOdds!, 0) / ms2_1_withOdds.length;
-    const avgDrawOdd = ms2_1_withOdds.filter(m => m.msDrawOdds).reduce((sum, m) => sum + m.msDrawOdds!, 0) / ms2_1_withOdds.filter(m => m.msDrawOdds).length || 0;
-
-    console.log(`\n   2/1 Matches - Average MS Odds:`);
-    console.log(`   Home Win: ${avgHomeOdd.toFixed(2)} (HOME won FT!)`);
-    console.log(`   Draw: ${avgDrawOdd.toFixed(2)}`);
-    console.log(`   Away Win: ${avgAwayOdd.toFixed(2)} (Away was WINNING at HT!)`);
-
-    let favoriteWon = 0;
-    let underdogWon = 0;
-
-    for (const m of ms2_1_withOdds) {
-      if (m.msAwayOdds! < m.msHomeOdds!) {
-        // Away was favorite at HT, but home won = UNDERDOG
-        underdogWon++;
-      } else if (m.msHomeOdds! < m.msAwayOdds!) {
-        // Home was favorite and won = FAVORITE
-        favoriteWon++;
-      }
-    }
-
-    console.log(`\n   2/1 - Who was favored vs who won:`);
-    console.log(`   Favorite won (Home was fav): ${favoriteWon} (${(favoriteWon / ms2_1_withOdds.length * 100).toFixed(1)}%)`);
-    console.log(`   Underdog won (Away was fav): ${underdogWon} (${(underdogWon / ms2_1_withOdds.length * 100).toFixed(1)}%) ⚠️`);
-  }
-
-  // ANALYSIS 3: Suspicious patterns
-  console.log('\n📈 ANALYSIS 3: SUSPICIOUS PATTERNS');
-  console.log('-'.repeat(80));
-
-  // Pattern 1: Heavy favorite loses after leading (1/2 with low home odds)
-  const suspicious_1_2 = ms1_2_withOdds.filter(m => m.msHomeOdds! < 1.5);
-  console.log(`\n⚠️  PATTERN 1: Heavy Home Favorite loses after HT lead (MS Home Odds < 1.5):`);
-  console.log(`   Count: ${suspicious_1_2.length}`);
-  if (suspicious_1_2.length > 0) {
-    const avgOdd = suspicious_1_2.reduce((sum, m) => sum + m.msHomeOdds!, 0) / suspicious_1_2.length;
-    console.log(`   Avg Home Odds: ${avgOdd.toFixed(2)}`);
-    console.log(`   Sample matches:`);
-    suspicious_1_2.slice(0, 5).forEach(m => {
-      console.log(`     ${m.league}: ${m.homeTeam} (${m.msHomeOdds}) vs ${m.awayTeam} (${m.msAwayOdds}) => HT: ${m.htHome}-${m.htAway}, FT: ${m.ftHome}-${m.ftAway}`);
-    });
-  }
-
-  // Pattern 2: Heavy away favorite loses after leading (2/1 with low away odds)
-  const suspicious_2_1 = ms2_1_withOdds.filter(m => m.msAwayOdds! < 1.5);
-  console.log(`\n⚠️  PATTERN 2: Heavy Away Favorite loses after HT lead (MS Away Odds < 1.5):`);
-  console.log(`   Count: ${suspicious_2_1.length}`);
-  if (suspicious_2_1.length > 0) {
-    const avgOdd = suspicious_2_1.reduce((sum, m) => sum + m.msAwayOdds!, 0) / suspicious_2_1.length;
-    console.log(`   Avg Away Odds: ${avgOdd.toFixed(2)}`);
-    console.log(`   Sample matches:`);
-    suspicious_2_1.slice(0, 5).forEach(m => {
-      console.log(`     ${m.league}: ${m.homeTeam} (${m.msHomeOdds}) vs ${m.awayTeam} (${m.msAwayOdds}) => HT: ${m.htHome}-${m.htAway}, FT: ${m.ftHome}-${m.ftAway}`);
-    });
-  }
-
-  // ANALYSIS 4: HT Odds vs MS Odds correlation
-  console.log('\n📈 ANALYSIS 4: HT ODDS CORRELATION');
-  console.log('-'.repeat(80));
-
-  const withHTOdds = reversalMatches.filter(m => m.htHomeOdds && m.htAwayOdds);
-  if (withHTOdds.length > 0) {
-    console.log(`\n   Matches with HT odds: ${withHTOdds.length}`);
-
-    let htCorrectlyPredicted = 0;
-    for (const m of withHTOdds) {
-      const htFav = m.htHomeOdds! < m.htAwayOdds! ? '1' : m.htAwayOdds! < m.htHomeOdds! ? '2' : 'X';
-      const htActual = m.htHome > m.htAway ? '1' : m.htAway > m.htHome ? '2' : 'X';
-      if (htFav === htActual) htCorrectlyPredicted++;
-    }
-    console.log(`   HT Favorite correctly led at HT: ${htCorrectlyPredicted}/${withHTOdds.length} (${(htCorrectlyPredicted / withHTOdds.length * 100).toFixed(1)}%)`);
-
-    // How often did HT favorite lose FT?
-    let htFavoriteLostFT = 0;
-    for (const m of withHTOdds) {
-      const htFav = m.htHomeOdds! < m.htAwayOdds! ? '1' : m.htAwayOdds! < m.htHomeOdds! ? '2' : 'X';
-      const ftActual = m.ftHome > m.ftAway ? '1' : m.ftAway > m.ftHome ? '2' : 'X';
-      if (htFav !== ftActual) htFavoriteLostFT++;
-    }
-    console.log(`   HT Favorite lost FT: ${htFavoriteLostFT}/${withHTOdds.length} (${(htFavoriteLostFT / withHTOdds.length * 100).toFixed(1)}%) ⚠️`);
-  }
-
-  // ANALYSIS 5: Score patterns
-  console.log('\n📈 ANALYSIS 5: SCORE PATTERNS IN REVERSALS');
-  console.log('-'.repeat(80));
-
-  // HT score distribution for reversals
-  const htScores: Record<string, number> = {};
-  for (const m of reversalMatches) {
-    const key = `${m.htHome}-${m.htAway}`;
-    htScores[key] = (htScores[key] || 0) + 1;
-  }
-
-  console.log('\nMost common HT scores in reversal matches:');
-  Object.entries(htScores)
-    .sort((a, b) => b[1] - a[1])
-    .slice(0, 10)
-    .forEach(([score, count]) => {
-      console.log(`   HT ${score}: ${count} matches`);
-    });
-
-  // FT score distribution
-  const ftScores: Record<string, number> = {};
-  for (const m of reversalMatches) {
-    const key = `${m.ftHome}-${m.ftAway}`;
-    ftScores[key] = (ftScores[key] || 0) + 1;
-  }
-
-  console.log('\nMost common FT scores in reversal matches:');
-  Object.entries(ftScores)
-    .sort((a, b) => b[1] - a[1])
-    .slice(0, 10)
-    .forEach(([score, count]) => {
-      console.log(`   FT ${score}: ${count} matches`);
-    });
-
-  // ANALYSIS 6: Goal difference patterns
-  console.log('\n📈 ANALYSIS 6: COMEBACK MAGNITUDE');
-  console.log('-'.repeat(80));
-
-  let comebackBy1 = 0;
-  let comebackBy2 = 0;
-  let comebackBy3Plus = 0;
-
-  for (const m of reversalMatches) {
-    const htDiff = Math.abs(m.htHome - m.htAway);
-    const ftDiff = Math.abs(m.ftHome - m.ftAway);
-
-    if (m.htft === '1/2') {
-      // Home was leading, away won
-      const margin = (m.ftAway - m.ftHome);
-      if (margin === 1) comebackBy1++;
-      else if (margin === 2) comebackBy2++;
-      else comebackBy3Plus++;
-    } else {
-      // Away was leading, home won
-      const margin = (m.ftHome - m.ftAway);
-      if (margin === 1) comebackBy1++;
-      else if (margin === 2) comebackBy2++;
-      else comebackBy3Plus++;
-    }
-  }
-
-  console.log(`\n   Comeback by 1 goal: ${comebackBy1} (${(comebackBy1 / reversalMatches.length * 100).toFixed(1)}%)`);
-  console.log(`   Comeback by 2 goals: ${comebackBy2} (${(comebackBy2 / reversalMatches.length * 100).toFixed(1)}%)`);
-  console.log(`   Comeback by 3+ goals: ${comebackBy3Plus} (${(comebackBy3Plus / reversalMatches.length * 100).toFixed(1)}%) ⚠️`);
-
-  // Show extreme comebacks
-  const extremeComebacks = reversalMatches
-    .filter(m => {
-      if (m.htft === '1/2') return (m.ftAway - m.ftHome) >= 2;
-      return (m.ftHome - m.ftAway) >= 2;
-    })
-    .sort((a, b) => {
-      const diffA = a.htft === '1/2' ? (a.ftAway - a.ftHome) : (a.ftHome - a.ftAway);
-      const diffB = b.htft === '1/2' ? (b.ftAway - b.ftHome) : (b.ftHome - b.ftAway);
-      return diffB - diffA;
-    })
-    .slice(0, 10);
-
-  console.log('\nTop 10 most extreme comebacks:');
-  extremeComebacks.forEach(m => {
-    const diff = m.htft === '1/2' ? (m.ftAway - m.ftHome) : (m.ftHome - m.ftAway);
-    console.log(`   ${m.league}: ${m.homeTeam} vs ${m.awayTeam} | HT: ${m.htHome}-${m.htAway} => FT: ${m.ftHome}-${m.ftAway} (Diff: ${diff})`);
-  });
-
-  console.log('\n' + '='.repeat(80));
-  console.log('✅ ANALYSIS COMPLETE');
-  console.log('='.repeat(80));
-
-  await prisma.$disconnect();
-}
-
-analyzeReversalMatches().catch(console.error);

check-lineup-data.ts

@@ -1,77 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-
-dotenv.config();
-
-// BigInt serialization fix
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-const matchId = '7cnm7h7qbsq2bbaxngusojh90';
-
-async function checkLineupData() {
-  const match = await prisma.liveMatch.findUnique({
-    where: { id: matchId },
-  });
-
-  if (!match) {
-    console.log('❌ Match not found');
-    return;
-  }
-
-  console.log('\n📊 LINEUP DATA INSPECTION');
-  console.log('='.repeat(80));
-  
-  console.log(`\n1. lineups field:`);
-  console.log(`   Type: ${typeof match.lineups}`);
-  console.log(`   Is null: ${match.lineups === null}`);
-  console.log(`   Content:`, JSON.stringify(match.lineups, null, 2));
-
-  console.log(`\n2. sidelined field:`);
-  console.log(`   Type: ${typeof match.sidelined}`);
-  console.log(`   Is null: ${match.sidelined === null}`);
-  console.log(`   Content:`, JSON.stringify(match.sidelined, null, 2));
-
-  console.log(`\n3. odds field:`);
-  console.log(`   Type: ${typeof match.odds}`);
-  console.log(`   Is null: ${match.odds === null}`);
-  
-  // Check if it's JSON object or string
-  if (match.odds) {
-    const oddsStr = typeof match.odds === 'string' ? match.odds : JSON.stringify(match.odds);
-    console.log(`   Length: ${oddsStr.length}`);
-    console.log(`   Preview: ${oddsStr.substring(0, 200)}...`);
-  }
-
-  console.log(`\n4. refereeName:`);
-  console.log(`   Value: ${match.refereeName}`);
-
-  // Now check what AI Engine sees
-  console.log('\n\n🔍 AI ENGINE PERSPECTIVE');
-  console.log('='.repeat(80));
-  
-  // Simulate AI Engine's lineup parsing
-  const lineups = match.lineups as any;
-  let homePlayers: any[] = [];
-  let awayPlayers: any[] = [];
-  
-  if (lineups && typeof lineups === 'object') {
-    if (lineups.home?.xi) {
-      homePlayers = lineups.home.xi;
-    }
-    if (lineups.away?.xi) {
-      awayPlayers = lineups.away.xi;
-    }
-  }
-
-  console.log(`\nHome lineup count: ${homePlayers.length}`);
-  console.log(`Away lineup count: ${awayPlayers.length}`);
-  console.log(`Lineup source would be: ${homePlayers.length >= 9 && awayPlayers.length >= 9 ? 'confirmed_live' : 'none/probable'}`);
-
-  await prisma.$disconnect();
-}
-
-checkLineupData().catch(console.error);

copy_key.exp

@@ -1,10 +0,0 @@
-#!/usr/bin/expect -f
-spawn ssh -p 2222 -o StrictHostKeyChecking=accept-new haruncan@95.70.252.214 "mkdir -p ~/.ssh && echo 'ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIGo7pRd2fozEvxIultfwgoajgNOzc0RVywcqrqgZho62 piton@Pitons-MacBook-Air.local' >> ~/.ssh/authorized_keys && chmod 700 ~/.ssh && chmod 600 ~/.ssh/authorized_keys"
-
-expect {
-    "assword:" {
-        send "M594xH%\$iM&4MM\r"
-        exp_continue
-    }
-    eof
-}

docker_info.txt

@@ -1,62 +0,0 @@
-Thu Apr 16 12:20:54 UTC 2026
-==== DOCKER PS ====
-CONTAINER ID   IMAGE                                 COMMAND                  CREATED         STATUS                          PORTS                                                                                NAMES
-78aab6872b85   gitea/runner-images:ubuntu-latest     "/bin/sleep 10800"       2 seconds ago   Up 1 second                                                                                                          GITEA-ACTIONS-TASK-185_WORKFLOW-Check-Docker-Pi_JOB-check-docker
-784ca4842e79   iddaai-be:latest                      "docker-entrypoint.s…"   6 minutes ago   Up 6 minutes                    3000/tcp, 127.0.0.1:1810->3005/tcp                                                   iddaai-be
-48f495d45025   iddaai-fe:latest                      "docker-entrypoint.s…"   2 hours ago     Up 2 hours                      127.0.0.1:1510->3000/tcp                                                             iddaai-fe
-a60b07c52d7a   gitea/act_runner:latest               "/sbin/tini -- run.sh"   22 hours ago    Up 22 hours                                                                                                          gitea_runner
-436552af4199   iddaai-ai-engine                      "uvicorn main:app --…"   23 hours ago    Up 23 hours (healthy)           8000/tcp                                                                             iddaai-ai-engine
-696050fc89de   postgres:17-alpine                    "docker-entrypoint.s…"   23 hours ago    Up 23 hours (healthy)           5432/tcp                                                                             iddaai-postgres
-abcc43242dbb   redis:7-alpine                        "docker-entrypoint.s…"   23 hours ago    Up 23 hours (healthy)           6379/tcp                                                                             iddaai-redis
-da0f2d5bc898   temporalio/auto-setup:latest          "/etc/temporal/entry…"   3 weeks ago     Up 8 days                       6933-6935/tcp, 6939/tcp, 7233-7235/tcp, 7239/tcp                                     temporal
-4768eec66926   ghcr.io/gitroomhq/postiz-app:latest   "docker-entrypoint.s…"   3 weeks ago     Up 8 days                       0.0.0.0:4007->5000/tcp, [::]:4007->5000/tcp                                          postiz
-5cfb55782d8b   postgres:16                           "docker-entrypoint.s…"   3 weeks ago     Up 8 days                       5432/tcp                                                                             temporal-postgresql
-cf8591458662   redis:7.2                             "docker-entrypoint.s…"   3 weeks ago     Up 8 days (healthy)             6379/tcp                                                                             postiz-redis
-0108dc0b875d   postgres:17-alpine                    "docker-entrypoint.s…"   3 weeks ago     Up 8 days (healthy)             5432/tcp                                                                             postiz-postgres
-c88a569ddb22   elasticsearch:8.16.2                  "/bin/tini -- /usr/l…"   3 weeks ago     Up 8 days                       9200/tcp, 9300/tcp                                                                   temporal-elasticsearch
-208bbf92c2d8   temporalio/ui:latest                  "./start-ui-server.sh"   3 weeks ago     Up 8 days                       0.0.0.0:8085->8080/tcp, [::]:8085->8080/tcp                                          temporal-ui
-a0555f255857   haruncan-studio-fe:latest             "/docker-entrypoint.…"   3 weeks ago     Up 8 days                       0.0.0.0:1509->80/tcp, [::]:1509->80/tcp                                              haruncan-studio-fe-container
-7591abf68bf5   backend-haruncan-studio:latest        "docker-entrypoint.s…"   3 weeks ago     Up 8 days                       0.0.0.0:1809->3000/tcp, [::]:1809->3000/tcp                                          backend-haruncan-studio-container
-96d02609b108   ui-indir:latest                       "docker-entrypoint.s…"   5 weeks ago     Up 8 days                       0.0.0.0:1507->3000/tcp, [::]:1507->3000/tcp                                          ui-indir-container
-f67335b1625f   ghcr.io/open-webui/open-webui:main    "bash start.sh"          6 weeks ago     Up 8 days (healthy)             0.0.0.0:3001->8080/tcp, [::]:3001->8080/tcp                                          openclaw
-24b3c6e32817   gitea/gitea:latest                    "/usr/bin/entrypoint…"   6 weeks ago     Up 8 days                       0.0.0.0:222->22/tcp, [::]:222->22/tcp, 0.0.0.0:1224->3000/tcp, [::]:1224->3000/tcp   gitea
-4e64e3199178   postgres:14                           "docker-entrypoint.s…"   6 weeks ago     Up 8 days                       5432/tcp                                                                             gitea_db
-cb7fdcbcd79f   postgres:16-alpine                    "docker-entrypoint.s…"   6 weeks ago     Up 8 days                       5432/tcp                                                                             backend_db
-f0784aedcadf   redis:alpine                          "docker-entrypoint.s…"   6 weeks ago     Up 8 days                       6379/tcp                                                                             apps_redis
-fdc89d4a236a   portainer/portainer-ce:latest         "/portainer"             6 weeks ago     Up 8 days                       8000/tcp, 9443/tcp, 0.0.0.0:9000->9000/tcp, [::]:9000->9000/tcp                      portainer
-2de41ca39c1f   backend-proje:latest                  "docker-entrypoint.s…"   2 months ago    Restarting (1) 35 seconds ago                                                                                        backend-container
-89268da2ab86   skript-ui                             "docker-entrypoint.s…"   2 months ago    Up 8 days                       0.0.0.0:1506->3000/tcp, [::]:1506->3000/tcp                                          ui-skript-container
-8fced773c984   skript-be                             "docker-entrypoint.s…"   2 months ago    Exited (1) 8 days ago                                                                                                backend-skript-container
-ec90982f14b6   backend-digicraft                     "docker-entrypoint.s…"   2 months ago    Up 8 days                       0.0.0.0:1805->3001/tcp, [::]:1805->3001/tcp                                          backend-digicraft-container
-4eec58a7f453   ui-digicraft                          "/docker-entrypoint.…"   2 months ago    Up 8 days                       0.0.0.0:1505->80/tcp, [::]:1505->80/tcp                                              ui-digicraft-container
-37f844a6cd20   frontend-proje:latest                 "docker-entrypoint.s…"   2 months ago    Up 8 days                       0.0.0.0:1800->3000/tcp, [::]:1800->3000/tcp                                          frontend-container
-==== DOCKER STATS ====
-CONTAINER ID   NAME                                                               CPU %     MEM USAGE / LIMIT   MEM %     NET I/O           BLOCK I/O     PIDS
-78aab6872b85   GITEA-ACTIONS-TASK-185_WORKFLOW-Check-Docker-Pi_JOB-check-docker   1.07%     0B / 0B             0.00%     1.12MB / 13.6kB   135kB / 0B    12
-784ca4842e79   iddaai-be                                                          0.00%     0B / 0B             0.00%     12.9MB / 1.04MB   250kB / 0B    18
-48f495d45025   iddaai-fe                                                          0.00%     0B / 0B             0.00%     491kB / 221kB     1.84MB / 0B   26
-a60b07c52d7a   gitea_runner                                                       0.10%     0B / 0B             0.00%     25.6MB / 24.8MB   4MB / 0B      11
-436552af4199   iddaai-ai-engine                                                   0.14%     0B / 0B             0.00%     881kB / 840kB     175MB / 0B    20
-696050fc89de   iddaai-postgres                                                    0.00%     0B / 0B             0.00%     130MB / 311MB     685MB / 0B    13
-abcc43242dbb   iddaai-redis                                                       2.39%     0B / 0B             0.00%     222kB / 126B      3.95MB / 0B   6
-da0f2d5bc898   temporal                                                           1.49%     0B / 0B             0.00%     1.75GB / 1.88GB   300MB / 0B    15
-4768eec66926   postiz                                                             0.54%     0B / 0B             0.00%     8.09MB / 4.33MB   474MB / 0B    152
-5cfb55782d8b   temporal-postgresql                                                0.03%     0B / 0B             0.00%     1.88GB / 1.75GB   57.1MB / 0B   39
-cf8591458662   postiz-redis                                                       0.17%     0B / 0B             0.00%     1.17MB / 545kB    23.4MB / 0B   6
-0108dc0b875d   postiz-postgres                                                    0.00%     0B / 0B             0.00%     945kB / 176kB     46.1MB / 0B   8
-c88a569ddb22   temporal-elasticsearch                                             0.18%     0B / 0B             0.00%     763kB / 96.1kB    288MB / 0B    94
-208bbf92c2d8   temporal-ui                                                        0.00%     0B / 0B             0.00%     655kB / 22.7kB    66.6MB / 0B   8
-a0555f255857   haruncan-studio-fe-container                                       0.00%     0B / 0B             0.00%     2.13MB / 7.98MB   4.78MB / 0B   5
-7591abf68bf5   backend-haruncan-studio-container                                  0.00%     0B / 0B             0.00%     776kB / 724kB     127MB / 0B    17
-96d02609b108   ui-indir-container                                                 0.00%     0B / 0B             0.00%     118MB / 27.3MB    139MB / 0B    11
-f67335b1625f   openclaw                                                           0.11%     0B / 0B             0.00%     652kB / 16.4kB    1GB / 0B      19
-24b3c6e32817   gitea                                                              2.44%     0B / 0B             0.00%     1.69GB / 1.24GB   200MB / 0B    20
-4e64e3199178   gitea_db                                                           0.74%     0B / 0B             0.00%     1.03GB / 1.25GB   69.1MB / 0B   10
-cb7fdcbcd79f   backend_db                                                         0.00%     0B / 0B             0.00%     677kB / 126B      41.4MB / 0B   6
-f0784aedcadf   apps_redis                                                         0.23%     0B / 0B             0.00%     677kB / 126B      31MB / 0B     6
-fdc89d4a236a   portainer                                                          0.00%     0B / 0B             0.00%     4.08MB / 20.8MB   126MB / 0B    7
-2de41ca39c1f   backend-container                                                  0.00%     0B / 0B             0.00%     0B / 0B           0B / 0B       0
-89268da2ab86   ui-skript-container                                                0.00%     0B / 0B             0.00%     1.71MB / 11.3kB   51.3MB / 0B   11
-ec90982f14b6   backend-digicraft-container                                        0.06%     0B / 0B             0.00%     2.41MB / 436kB    142MB / 0B    38
-4eec58a7f453   ui-digicraft-container                                             0.00%     0B / 0B             0.00%     3.95MB / 27.3MB   5.8MB / 0B    5
-37f844a6cd20   frontend-container                                                 0.01%     0B / 0B             0.00%     1.93MB / 3.11MB   13.6MB / 0B   11

lint-output.txt

@@ -1,4 +0,0 @@
-
-> Suggest-Bet-BE@0.0.1 lint
-> eslint "{src,apps,libs,test}/**/*.ts" --fix
-

mds/SESSION_HANDOFF.md

@@ -0,0 +1,305 @@
+# SESSION HANDOFF — iddaai sistem durumu
+
+**Son güncelleme**: 2026-05-25 ~23:00 (Windows'tan Mac'e geçiş öncesi)
+**Hedef**: Başka makinede / yeni Claude session'ında bu doc tek başına okunup işin nerede kaldığı anlaşılabilmeli.
+
+---
+
+## 🚨 EN SON DURUM (Mac'e geçmeden önce oku)
+
+### Validation backtest ÖLDÜ
+- Pencere: 2026-05-01 → 2026-05-14, 1500 maç
+- **1200/1500'de SSH tunnel düşünce process sessizce öldü**
+- **CSV kayıp** — script eski versiyondu, sadece sonda yazıyordu
+- Sebep: localhost:5432 erişimi kayboldu, psycopg2 connection error
+
+### Script DÜZELTILDI (Mac'te kullanılabilir)
+`scripts/diagnostic_backtest.py` artık **crash-safe**:
+- `--checkpoint-every 100` → her 100 maçta partial CSV diske yazılır
+- Crash sonrası tekrar koşulunca **otomatik kaldığı yerden devam**
+- Checkpoint dosyası: `reports/_checkpoint_<window-key>.csv`
+- `--no-resume` flag fresh başlamak için
+
+### Git push BEKLİYOR
+- 36 dosya **commit edildi (local)** — bkz "Bu seansta yapılan KOD değişiklikleri"
+- Push **auth hatası** verdi (gitea credentials cached değil)
+- **User Mac'te push yapacak** (Gitea Personal Access Token gerekli, repo write yetkisi)
+
+### Mac'te yapılacaklar (öncelikli sırayla)
+1. Repo'yu clone et veya OneDrive'dan kopyala (eğer Mac OneDrive senkronize ediyorsa)
+2. `git push origin main` ile pending commit'i remote'a yolla
+3. SSH tunnel kur (Pi @ 95.70.252.214, port 2222) → DB için tunnel localhost:5432
+4. Yeni Claude session'ı başlat, bu dosyayı oku, devam et
+5. Backtest tekrar koştur (çoktan eski versiyondu, şimdi crash-safe)
+   ```bash
+   cd ai-engine
+   export DATABASE_URL="postgresql://iddaai_user:IddaA1_S4crET!@localhost:5432/iddaai_db?schema=public"
+   export PYTHONIOENCODING=utf-8
+   python scripts/diagnostic_backtest.py --start 2026-05-01 --end 2026-05-14 --max-matches 1500
+   # ölürse, aynı komutu tekrar koş — checkpoint'ten devam eder
+   ```
+
+### Mevcut sağlam veri
+Validation kayıp ama elimizde **in-sample backtest** ve **grid search** çıktıları var:
+- `reports/diagnostic_backtest_20260525_035649.{csv,json,txt}` — 1000 maç, May 11-24
+- `reports/filter_optimization_patch.json` — grid search winners
+- Bu data ile in-sample analiz tamamlandı, validation eksik
+
+---
+
+---
+
+## 🎯 Üst-seviye hedef
+
+Sistem **maç başı-1 saat** kullanıcı tetiklemesiyle çalışacak. Bahis uzmanı seviyesinde:
+- **main_pick + value_pick** (sistemin önerdiği)
+- **Tüm market olasılıkları** (MS, HT, OU05-45, BTTS, OE, DC, HTFT, HCAP, Cards, Corners)
+- **Net HT + FT skoru** + **Top-5 olası skor dağılımı**
+- **Evidence panel**: lineup impact, son 5 maç, h2h, hakem profili, benzer-oran-band geçmişi
+
+Ürün modeli: hem user kendi bahisini oynar, hem sistem para kazanırsa abonelik satılır.
+Hedef ROI: **≥%10**. Günde **3-5 kaliteli bahis**.
+
+Detaylı requirements doc: bu dosyanın altında, "Requirements Spec" bölümü.
+
+---
+
+## 🟢 Şu an arka planda KOŞAN işler
+
+### 1. Validation backtest (LOCAL — bu laptop)
+- **Script**: `ai-engine/scripts/diagnostic_backtest.py`
+- **Komut**: `python scripts/diagnostic_backtest.py --start 2026-05-01 --end 2026-05-14 --max-matches 1500`
+- **Log**: `ai-engine/validation_full.log` (OneDrive senkronize)
+- **Çıkış**: bittiğinde `ai-engine/reports/diagnostic_backtest_<timestamp>.{csv,json,txt}`
+- **Tahmini bitiş**: 2026-05-25 ~22:00 (yaklaşık)
+- **Amaç**: Yeni kodla (calibrator + ev_edge veto + envelope + coherence + BTTS mute) **out-of-sample** doğrulama
+- **Risk**: Laptop uyursa ölür. Bitmesini beklemen lazım VEYA partial sonuçla devam.
+
+```powershell
+# Status check (kendin)
+$log='C:\Users\fahri\OneDrive\المستندات\GitHub\iddaai\iddaai-be\ai-engine\validation_full.log'
+Select-String $log 'rate=|Outputs:' | Select-Object -Last 3 | ForEach-Object {$_.Line}
+```
+
+### 2. Feeder historical scan (REMOTE — Pi server)
+- **Konum**: SSH @ haruncan@95.70.252.214:2222 → docker container `iddaai-be` → pm2
+- **PM2 process**: `feeder-historical` (id=1)
+- **Log rotation**: pm2-logrotate kurulu (max 30MB/dosya, 3 dosya, gzip)
+- **Davranış**: 2026-05-03'ten geriye 2023-06-01'e kadar mackolik'ten odds/lineup patch
+- **Otomatik restart**: 502 olunca 30 sn delay sonra restart (max 1000 kez)
+- **Beklenen süre**: 24-72 saat
+
+```bash
+# Status (kendin SSH'la)
+sudo docker exec iddaai-be pm2 list
+sudo docker exec iddaai-be pm2 logs feeder-historical --lines 30 --nostream
+```
+
+---
+
+## 📝 Bu seansta yapılan KOD değişiklikleri
+
+Hepsi local repo'da, OneDrive senkronize edecek, başka makinede pull etmesen de açtığında orada olacak.
+
+### A. Settlement / data layer
+| Dosya | Değişiklik |
+|---|---|
+| `iddaai-be/prisma.config.ts` | `.env` fallback ekledim (`.env.local` üstüne) — `prisma generate` çalışsın diye |
+| `iddaai-be/src/tasks/prediction-settlement.market-resolver.ts` | DC parser ayraçsız "1X/X2/12" kabul ediyor + HT_OU05/HT_OU15/HT_OU25 resolver eklendi |
+| `iddaai-be/src/tasks/feature-enrichment.task.ts` **(YENİ)** | Cron 08:15 — eksik football_ai_features row insert + odds_movement SQL backfill |
+| `iddaai-be/src/tasks/python-enrichment.task.ts` **(YENİ)** | Cron 08:25 — Python `enrich_ai_features.py` subprocess |
+| `iddaai-be/src/tasks/tasks.module.ts` | İki yeni task register |
+| `iddaai-be/src/scripts/run-feature-enrichment.ts` **(YENİ)** | Manuel one-shot trigger |
+
+### B. AI engine — betting brain
+`iddaai-be/ai-engine/services/betting_brain.py` — büyük revizyon:
+- **HARD_MIN_SAMPLES = 50** floor (calibrator bypass <50 sample)
+- **`ev_edge < 0.0` HARD VETO** (`negative_ev_edge`)
+- **`ev_edge >= 0.20` HARD VETO** (`ev_edge_too_high_trap`)
+- **`MUTED_MARKETS = {"BTTS"}`** — backtest no profitable config bulduğu için
+- **`MARKET_OPTIMAL_FILTERS`** — MS ve OU25 için grid-search'ten gelen optimal envelope
+- **`_score_consistent_markets()`** — skor tahminine uymayan picks elimine
+- **`judge()` score coherence filter** — main_pick coherent set'ten seçilir
+- **HTFT reversal cross-check** — Man City 1/2 senaryosu
+
+### C. AI engine — model & calibration
+| Dosya | Değişiklik |
+|---|---|
+| `ai-engine/models/calibration.py` | HARD_MIN_SAMPLES floor + sample-weighted blend formülü değişti |
+| `ai-engine/models/calibration/*.pkl` | **10 calibrator retrain** (ms_home/draw/away, ou15/25/35, btts, ht_home/draw/away) — 4989-5000 sample her biri |
+
+### D. AI engine — orchestrator feature builder
+`ai-engine/services/orchestrator/feature_builder.py`:
+- Hardcoded `home_position=10, away_position=10` → real `data.home_position` kullanılıyor
+- Cup detection upper'a taşındı, `is_cup_match` UpsetEngine'e geçiyor
+- Total teams parametresi UpsetEngine'e geçiyor
+
+`ai-engine/services/orchestrator/data_loader.py`:
+- `_estimate_league_position` artık **sezon filtresi** (son 300 gün) kullanıyor
+
+### E. AI engine — scripts (yeni)
+| Dosya | Ne yapıyor |
+|---|---|
+| `ai-engine/scripts/diagnostic_backtest.py` | Per-bet diagnostic backtest (CSV+JSON+TXT output) |
+| `ai-engine/scripts/analyze_backtest_csv.py` | Backtest CSV üzerinde root-cause hipotez testleri |
+| `ai-engine/scripts/optimize_filters.py` | Grid search per-market optimal threshold |
+| `ai-engine/scripts/compare_backtests.py` | İki CSV karşılaştırması verdict ile |
+| `ai-engine/scripts/test_score_coherence.py` | Coherence filter smoke test (LAFC senaryosu) |
+
+### F. Social poster modülü (NestJS)
+| Dosya | Değişiklik |
+|---|---|
+| `src/modules/social-poster/social-poster.service.ts` | Cron 15→10 dk, window 10-60, MAX_POSTS_PER_RUN, getHealthStatus() |
+| `src/modules/social-poster/image-renderer.service.ts` | SEO filename + metadata sidecar (.json) |
+| `src/modules/social-poster/caption-generator.service.ts` | SEO hashtag stratejisi (12 küratör tag) |
+| `src/modules/social-poster/social-poster.controller.ts` | `/health` public + `/preview-png/:matchId` + `/run-now` endpoints |
+| `mds/SOCIAL_POSTER_SETUP.md` **(YENİ)** | Env vars + API key alma adımları + test komutları |
+
+### G. Modern image rendering (deneme)
+| Dosya | Açıklama |
+|---|---|
+| `src/scripts/render-social-card-v3.ts` | satori + resvg-js ile modern HTML→PNG rendering (Twemoji top + bayrak) |
+| `src/modules/social-poster/assets/*.svg` | Twemoji futbol/basket/bayrak SVG'leri |
+
+### H. Yapılan DB değişiklikleri (idempotent — tekrar koşturulursa sorun yok)
+| İşlem | Etki |
+|---|---|
+| `football_ai_features` 4008+ satır backfill | Son 60 günün FT maçları için feature row var artık (calculator_ver=feature_enrichment_task_v1) |
+| Python enrichment koştu | h2h, referee, possession, league_avg, implied_* hepsi gerçek değerlerle dolu (181,614+ satır enriched) |
+| Calibrator dosyaları yazıldı | `ai-engine/models/calibration/*.pkl` overwritten |
+
+---
+
+## 📂 Önemli dosya konumları (OneDrive synced)
+
+```
+iddaai-be/
+├── mds/
+│   ├── SESSION_HANDOFF.md            ← BU DOSYA
+│   └── SOCIAL_POSTER_SETUP.md        ← social poster env+keys
+├── ai-engine/
+│   ├── reports/                      ← BACKTEST CIKTILARI
+│   │   ├── diagnostic_backtest_*.csv,json,txt
+│   │   └── filter_optimization_patch.json
+│   ├── validation_full.log           ← validation backtest canlı log
+│   ├── diagnostic_backtest_run.log   ← önceki backtest log
+│   ├── enrichment_run3.log           ← enrichment koşma log
+│   └── calibration_run.log           ← calibrator retrain log
+├── public/predictions/               ← render edilmiş social card PNG/JSON
+└── src/scripts/                      ← tüm yeni script'ler
+```
+
+---
+
+## 🔑 Erişim bilgileri
+
+### Pi sunucu (feeder + prod stack)
+- **SSH**: `haruncan@95.70.252.214:2222`
+- **Şifre**: `M594xH%$iM&4MM`
+- **Plink kullan**: `~/plink.exe -ssh -P 2222 -pw '<password>' -hostkey 'SHA256:iq0YVI/4J897sf9dkksI7QzetpLCD0l57ZMX4UissI8' haruncan@95.70.252.214`
+- **Docker**: `iddaai-be`, `iddaai-ai-engine`, `iddaai-fe`, `iddaai-postgres`, `iddaai-redis`, `gitea`
+
+### DB (uzak Postgres @ Pi)
+- **SSH tunnel function**: `iddaai-db` PowerShell fonksiyonu (yerel makinedeki profile'da kayıtlı)
+- **Tunnel: localhost:5432 → Pi:5432**
+- **Connection string**: `postgresql://iddaai_user:IddaA1_S4crET!@localhost:5432/iddaai_db?schema=public`
+- **MCP**: Claude'un postgres MCP'si bu tunnel üzerinden çalışıyor (restricted mode, read-only)
+
+---
+
+## 📊 BACKTEST sonuçları geçmişi
+
+### Backtest #1 — In-sample grid search (2026-05-11 → 05-24, 1000 maç)
+- **CSV**: `ai-engine/reports/diagnostic_backtest_20260525_035649.csv`
+- **TXT**: `ai-engine/reports/diagnostic_backtest_20260525_035649.txt`
+- **Toplam playable**: 524 bet
+- **Hit rate**: %54.77
+- **ROI**: **−%16.73** (baseline kötü)
+- **Grid-search'ten çıkan optimal filtreler (in-sample)**:
+  - MS: edge [-5%, +15%], V27 AGREE zorunlu → +%8.23 (21 bet)
+  - OU25: odds ≥ 1.80, edge ≤ +15% → +%28.91 (20 bet)
+  - BTTS: tüm config'lerde kayıp → MUTE
+- **Aggregate optimize**: 95 bet, ROI +%2.16 (in-sample)
+
+### Backtest #2 — Validation (2026-05-01 → 05-14, KOŞUYOR)
+- **Bitince konum**: `ai-engine/reports/diagnostic_backtest_<yeni_timestamp>.{csv,json,txt}`
+- **Karşılaştırma çalıştır**: `python scripts/compare_backtests.py` (otomatik en yeni 2'yi alır)
+- **Beklenen sonuç**: ROI ≥ 0 → out-of-sample doğrulama BAŞARILI; in-sample overfit değil
+
+---
+
+## ❓ Backtest BİTTİĞİNDE yapılacak (yeni session'da bu kısımdan başla)
+
+### 1. Sonucu oku
+```powershell
+cd C:\Users\fahri\OneDrive\المستندات\GitHub\iddaai\iddaai-be\ai-engine
+Get-ChildItem reports\diagnostic_backtest_*.txt | Sort-Object LastWriteTime -Descending | Select-Object -First 1 | Get-Content
+```
+
+### 2. Karşılaştır
+```powershell
+python scripts\compare_backtests.py
+```
+
+Bu otomatik en yeni 2 backtest'i karşılaştırır, **VERDICT** verir:
+- ✅ "FILTERS WORK" → ROI pozitif AND improved
+- 🟡 "PARTIAL" → improved ama hâlâ negatif
+- ❌ "OVERFITTING" → validation ROI collapse
+
+### 3. Karara göre 2 yol
+
+**Eğer ROI ≥ +%2 ve overfit yok:**
+- `/sc:design` ile UI/API contract → Sprint 1
+- Sprint 1: top-5 skor + evidence panel + "why" cümlesi
+- Test edip prod'a aç
+
+**Eğer ROI negatif veya overfit:**
+- `analyze_backtest_csv.py` ile loss diagnostic
+- Hangi market hâlâ kötü → tighten filter veya mute
+- Calibrator recalibrate (özellikle BTTS dışındakiler için yeni sample)
+- Tekrar backtest
+
+---
+
+## ⚠️ Bilinen açık problemler / sorular
+
+1. **Coherence filter validate edilmedi production-side** — smoke test 20/20 ama gerçek production data ile karşılaştırma yok
+2. **Lineup-overlap last-5 hesabı** — yazılmadı, requirements doc'ta F8 var
+3. **Skor top-5 distribution** — Poisson zaten hesaplıyor, surface edilmedi (UI tarafı)
+4. **"Why" cümlesi main_pick'te** — boş, doldurulması gerek
+5. **Cards/Corners/RED CARD model** — yok, "henüz desteklenmiyor" placeholder ile bırak (kullanıcı onayladı: mevcut market'ler sağlamlaşsın)
+6. **Orphan match_id 51 satır** — `prediction_runs` içinde, `matches`'ta yok. Sample noise, geçiştirilebilir.
+7. **opening_value feeder bug** — `odds_movement_*` SQL yazıyor ama tüm değerler 0 (opening == closing). Feeder upstream sorun. Düşük öncelik.
+
+---
+
+## 🚦 Yeni Claude session'ında ilk komut
+
+```
+Bu projeye yeni bağlandım. Lütfen aşağıdaki dosyayı oku ve bana proje durumunu özet ver:
+
+C:\Users\fahri\OneDrive\المستندات\GitHub\iddaai\iddaai-be\mds\SESSION_HANDOFF.md
+
+Sonra validation backtest'in sonucuna bak:
+- C:\Users\fahri\OneDrive\المستندات\GitHub\iddaai\iddaai-be\ai-engine\reports\
+  içindeki en yeni diagnostic_backtest_*.txt dosyasını oku
+- compare_backtests.py script'ini koş, verdict göster
+- Verdict'e göre sonraki adımı öner
+```
+
+Buradan devam eder. Tüm context bu doc'ta + dosyalarda + DB'de.
+
+---
+
+## 🛠️ Requirements spec (sıkıştırılmış)
+
+**Ürün**: UI-tetikli per-match analiz, bahis uzmanı seviyesi
+**Trigger**: User tıklar, on-demand
+**Output**: main_pick + value_pick + tüm market olasılıkları + tek HT/FT skoru + top-5 skor dağılımı + evidence panel
+**Kapsam**: Mevcut market'ler sağlamlaştırılır, yeni market eklenmez (kullanıcı onayı)
+**Quality bar**: Calibration sapması ±2-5pp per market, NaN yok, response <3sn
+**Validation**: Out-of-sample backtest (1500 maç, May 1-14) — KOŞUYOR
+
+---
+
+**SON NOT**: Backtest'in TAMAMLANMASINI bekle (~22:00). Laptop'u kapatma. Bittiğinde OneDrive senkronize eder, başka makinede otomatik orada olur. Yeni session'da bu dosyayı oku, sonuçlara bak, devam et.

mds/SOCIAL_POSTER_SETUP.md

@@ -0,0 +1,212 @@
+# Social Poster — Setup & Operations
+
+Otomatik tahmin kartı üretip Twitter / Facebook / Instagram'a postlayan modül.
+Cron her **10 dakikada bir** çalışır, **yaklaşan 10-60 dk içindeki maçları**
+yakalar, AI Engine'den tahmin alır, 1080×1080 görsel üretir, caption üretir,
+3 platforma post eder.
+
+## 1) ENV değişkenleri
+
+`.env`'ye ekle:
+
+```bash
+# Master switch — false ise cron çalışmaz, manual endpoint'ler de boş döner.
+SOCIAL_POSTER_ENABLED=true
+
+# Hangi sporlar (virgüllü liste). Varsayılan: football,basketball
+SOCIAL_POSTER_SPORTS=football,basketball
+
+# Yaklaşan maç penceresi (dakika). Varsayılan 10-60.
+SOCIAL_POSTER_WINDOW_MIN=10
+SOCIAL_POSTER_WINDOW_MAX=60
+
+# Tek cron koşusunda kaç maç post edilir (rate-limit koruması). Varsayılan 5.
+SOCIAL_POSTER_MAX_PER_RUN=5
+
+# Public base URL — Instagram media upload için fotoğrafın HTTPS'ten erişilebilir
+# olması ŞART. Localhost ile IG çalışmaz; production domain veya ngrok kullan.
+APP_BASE_URL=https://api.iddaai.com
+
+# AI Engine URL (orchestrator)
+AI_ENGINE_URL=http://localhost:8000
+
+# ─── Twitter / X ───
+TWITTER_API_KEY=...
+TWITTER_API_SECRET=...
+TWITTER_ACCESS_TOKEN=...
+TWITTER_ACCESS_SECRET=...
+
+# ─── Meta (Facebook + Instagram) ───
+META_PAGE_ACCESS_TOKEN=...   # FB Page'in long-lived access token'ı
+META_PAGE_ID=...             # FB Page numeric ID
+META_IG_USER_ID=...          # IG Business account numeric ID
+META_GRAPH_API_VERSION=v25.0 # opsiyonel
+
+# ─── Caption AI (opsiyonel — yoksa template caption kullanılır) ───
+ENABLE_GEMINI=true
+GEMINI_API_KEY=...
+GEMINI_MODEL=gemini-1.5-flash
+
+# Veya local Ollama:
+OLLAMA_BASE_URL=http://localhost:11434
+SOCIAL_POSTER_OLLAMA_MODEL=llama3.1
+```
+
+## 2) API anahtarlarını alma
+
+### Twitter / X
+1. https://developer.x.com → Project + App oluştur
+2. App'ın "Keys and tokens" → "API Key", "API Secret" al
+3. User authentication settings → "Read and write and Direct Message"
+4. "Access Token and Secret" generate et (bu hesap adına post eder)
+5. **Free tier**: 1500 tweet/ay, 50 post/24 saat — 10 dk'lık cron'la günde
+   ~144 koşu × 5 post = 720 potansiyel post → free tier yetmez, **Basic plan
+   ($200/ay)** lazım. Cron interval'i 30 dk'ya alıp 50/gün kalmak istersen
+   `@Cron("*/30 * * * *")` olarak değiştir.
+
+### Meta (Facebook + Instagram)
+1. https://developers.facebook.com → App oluştur (type: Business)
+2. Facebook Page bağla (mevcut sayfan yoksa oluştur)
+3. Instagram Business hesabını Facebook Page'e bağla
+4. Graph API Explorer'dan **page access token** al (User token değil!)
+5. Long-lived token'a çevir (60 gün geçerli, refresh edilebilir)
+6. **Page ID**: `https://graph.facebook.com/me/accounts?access_token=...`
+7. **IG User ID**: `graph.facebook.com/{pageId}?fields=instagram_business_account&access_token=...`
+8. Required permissions: `pages_show_list`, `pages_manage_posts`,
+   `pages_read_engagement`, `instagram_basic`, `instagram_content_publish`
+
+### Gemini (caption AI — opsiyonel)
+- https://aistudio.google.com → API key (free tier yeterli, günde ~1500 istek)
+- `ENABLE_GEMINI=true` + `GEMINI_API_KEY=...`
+- Gemini yoksa template caption kullanılır (yine SEO'lu, sadece daha statik)
+
+## 3) Test komutları
+
+```bash
+# Servisi başlat
+npm run start:dev
+
+# Health endpoint — auth gerekmez
+curl http://localhost:3005/social-poster/health | jq
+
+# Manuel preview (görsel + JSON) — superadmin token gerekir
+curl -H "Authorization: Bearer $TOKEN" \
+     http://localhost:3005/social-poster/preview/<matchId>
+
+# Görseli tarayıcıda direkt göster
+open http://localhost:3005/social-poster/preview-png/<matchId>?token=$TOKEN
+
+# Manuel post (tek maç, tüm platformlara) — superadmin token
+curl -X POST -H "Authorization: Bearer $TOKEN" \
+     http://localhost:3005/social-poster/post/<matchId>
+
+# Cron'u beklemeden full sweep koş — superadmin token
+curl -X POST -H "Authorization: Bearer $TOKEN" \
+     http://localhost:3005/social-poster/run-now
+```
+
+## 4) SEO özellikleri
+
+### Image dosya adı (SEO)
+Eskiden: `prediction_basketball_xyz12345_1716595200000.jpg` (opaque)
+Yeni: `sampiyonlar-ligi-unicaja-malaga-vs-aek-20260525.jpg` (Google indexable)
+
+### Yan dosya: metadata sidecar
+Her görsel için aynı dizinde `.json`:
+- `title`, `description`, `og:*`, `schema.org SportsEvent`, `picks[]`
+- Sayfada `<head>` Open Graph + Twitter Cards bu dosyadan beslenir
+- Schema.org markup zengin sonuç (Google rich snippet) sağlar
+
+### Caption (SEO + hashtags)
+Her post 12'ye kadar küratör hashtag içerir:
+- Marka: `#MaçTahmini #İddaa #BugünMaç`
+- Spor: `#Futbol #Basketbol #FutbolTahmin`
+- Lig: `#SüperLig #PremierLeague #ŞampiyonlarLigi #EuroLeague #NBA`
+- Bölge: `#Türkiye #İngiltere #İspanya`
+- Takım: `#Galatasaray #Fenerbahçe`
+- Gün: `#PazarTahmini #CumartesiTahmini`
+- Market: `#AltÜst #KGVar #ÇifteŞans #MaçSonucu #Handikap`
+
+LLM (Gemini) caption üretiyorsa hashtag'leri çıkarır; sistem kendi
+hashtag set'ini ekler. Tutarlı index için tek kaynak.
+
+## 5) İzleme
+
+```bash
+# Health endpoint — periyodik monitor
+curl http://localhost:3005/social-poster/health | jq
+
+# Sample output:
+{
+  "enabled": true,
+  "sports": ["football", "basketball"],
+  "window_min_minutes": 10,
+  "window_max_minutes": 60,
+  "max_posts_per_run": 5,
+  "top_leagues_loaded": 42,
+  "posted_match_count": 137,
+  "last_run_at": "2026-05-25T03:10:00.123Z",
+  "last_run_result": { "posted": 4, "skipped": 1, "errors": 0 },
+  "twitter_available": true,
+  "meta_facebook_available": true,
+  "meta_instagram_available": true,
+  "ai_engine_url": "http://localhost:8000",
+  "app_base_url": "https://api.iddaai.com"
+}
+```
+
+`posted_match_count` `storage/social-poster-posted.json`'dan okunur, son 500
+match ID hafızada — aynı maçı 2 kere post etmez.
+
+## 6) Rate limit ipuçları
+
+| Platform | Free limit | Tedbir |
+|---|---|---|
+| Twitter | 50 post/24 saat | `SOCIAL_POSTER_MAX_PER_RUN=2` + cron `*/30` → günde ~96 |
+| Facebook | ~200 post/saat (Page) | Default config rahat |
+| Instagram | 25 post/24 saat | `MAX_PER_RUN=1` + cron `*/60` → günde 24, sınırın hemen altında |
+
+IG en sıkı sınır — production için **IG ayrı cron'da daha seyrek post**
+yapılması önerilir (kod henüz tek cron, ileride ayrılabilir).
+
+## 7) Hangi maçlar seçilir?
+
+`top_leagues.json` dosyasındaki league_id'ler içinden:
+- Şu anda 10-60 dakika sonra başlayacak
+- Daha önce post edilmemiş
+- `sport: football, basketball` filtresi geçen
+
+`top_leagues.json` yoksa **tüm liglerden** maç seçer (hacmi yüksek tutar).
+Sadece premium ligler postlamak istersen dosyayı doldur.
+
+## 8) Görsel formatı
+
+- **Boyut**: 1080×1080 (Instagram square — Twitter da kabul ediyor)
+- **Format**: JPEG, quality 94
+- **Tema**: Sport'a göre değişir — football yeşil, basketball turuncu
+- **İçerik**: Lig logosu + ülke bayrağı, takım logoları + adları, HT skor,
+  FT skor, top 3 tahmin (confidence ile), risk badge
+
+Card layout `image-renderer.service.ts` içinde — value-pick yıldız ile
+işaretli, scenario top 3 listelenir, footer alt'ta tarih + brand.
+
+## 9) Sık sorular
+
+**Q: Görseller nereye yazılıyor?**
+`public/predictions/` (gitignored). ServeStatic ile `/predictions/<file>.jpg`
+URL'inden erişilir.
+
+**Q: Eski görseller temizleniyor mu?**
+Hayır — manuel temizlik gerekir. Cron eklemek istersen `LimitResetterTask`
+örneği var.
+
+**Q: AI Engine çalışmıyorsa ne olur?**
+Cron tahmin alamaz, log'a hata düşer, devam eder. Sonraki koşuda dener.
+
+**Q: Bir maç 2 kere post ediliyor mu?**
+Hayır — `postedMatchIds` set'i Match ID bazında dedup yapar, dosyaya yazılır
+(restart-safe).
+
+**Q: Caption Gemini olmadan ne kadar iyi?**
+Template caption tüm bilgileri + 12 hashtag içerir. SEO açısından yeterli,
+sadece anlatım daha statik. Gemini ile her post için özgün metin.

mds/brainstorm-leagues-and-national-teams.md

@@ -0,0 +1,72 @@
+# Gereksinim Keşfi: Lig Etiketleme + Milli Takım / Dünya Kupası Desteği
+
+> `/sc:brainstorm` çıktısı — REQUIREMENTS ONLY. Tasarım/kod sonraki adım (/sc:design, /sc:workflow).
+> Tarih: 2026-06 · Kaynak: 10k backtest + canlı DB/API kanıtı.
+
+## Doğrulanmış Gerçekler (kanıta dayalı, varsayım değil)
+
+### Lig performansı
+- 10k backtest 18 ligde BET üretti; ROI dağılımı: ~7 güçlü kârlı (+25%..+102%),
+  ~4 başabaş, ~7 zararlı (−12%..−62%).
+- `live_matches` distinct lig ≠ `qualified_leagues.json` (48 lig). live_matches'te
+  qualified olmayan ligler var → kullanıcının "gereksiz ligler" sezgisi DOĞRU.
+- Lig isimleri backtest CSV'de boş; DB'den `leagues` tablosundan çözülür.
+
+### Milli takım / Dünya Kupası (ÖNEMLİ — ilk hipotez ÇÜRÜDÜ)
+- Milli takımlar DB'de VAR: Türkiye(9s2kpeunkes0g17l95r3t91j6, elo 1675),
+  Almanya(3l2t2db0c5ow2f7s7bhr6mij4, 1689), Kolombiya(1692), Andorra(1243).
+  ELO + matches_played (30-37) MEVCUT.
+- Milli maç hacmi yüksek: DK Elemeler 645, Hazırlık Maçları Ülkeler 522,
+  Uluslar Ligi 148, Avrupa Şamp. Elemeleri 120 bitmiş maç. Ayrı ligler halinde.
+- football_ai_features milli maçlar için %98 üretiliyor (196/200) — kulüpten yüksek.
+- **KÖK SEBEP (canlı API ile kanıtlandı):** Sistem milli maçı tahmin EDİYOR
+  (MS olasılıkları + 9-10 market geliyor, data_quality MEDIUM 0.57-0.74). Sorun:
+  `betting_brain approved=0` — hiçbir market "oynanabilir" işaretlenmiyor. Ortak
+  flag `ai_features_inferred_from_history` → data_quality MEDIUM tavanı (0.74) +
+  lig qualified değil → brain eşikleri geçilemiyor. Yani "yetersiz veri" mesajı
+  aslında "brain güvenmiyor, BET yok" demek. Model/veri sorunu DEĞİL, gate/tuning sorunu.
+
+## Kullanıcı Kararları (bu oturumda alındı)
+- Lig filtresi: **"Hepsi görünsün ama etiketli"** (gizleme yok; güven rozeti: Yüksek/Orta/Düşük).
+- Milli takım: başta "ayrı model" istendi; veri görülünce yön = mevcut motoru milli
+  maçlara uyarlamak (ayrı ML modeli gereksiz — ELO+feature+geçmiş zaten var).
+- Öncelik: **önce lig etiketleme** (hazır veri), sonra milli takım.
+- Dünya Kupası: hazırlık maçlarında test edilebilmeli (yakın takvim baskısı).
+
+## Fonksiyonel Gereksinimler
+
+### A. Lig Güven Etiketleme
+- FR-A1: Her lig için backtest'e dayalı güven seviyesi (Yüksek/Orta/Düşük) hesaplanmalı
+  (metrik: BET ROI + örneklem sayısı; düşük örneklem = otomatik Düşük/Bilinmiyor).
+- FR-A2: live_matches'teki maçlar lig güven rozetiyle gösterilmeli (gizlenmeden).
+- FR-A3: Etiket kaynağı tek yerde (config/tablo) tutulmalı, backtest tazelendikçe güncellenebilmeli.
+- FR-A4: Forward-test (Model Performansı) verisi biriktikçe etiketler canlı sonuçla doğrulanmalı.
+
+### B. Milli Takım / Dünya Kupası Desteği
+- FR-B1: Milli maçlarda da BET önerisi çıkabilmeli (şu an approved=0).
+- FR-B2: `ai_features_inferred_from_history` flag'i olan milli maçlar için data_quality
+  tavanı / brain eşikleri milli-maça uygun kalibre edilmeli (kör gevşetme DEĞİL).
+- FR-B3: Milli maç ligleri (DK Elemeler, Hazırlık Maçları Ülkeler, Uluslar Ligi, Avrupa
+  Şamp., Dünya Kupası) "tanınan" kapsama alınmalı (qualified benzeri).
+- FR-B4: Hazırlık maçlarında uçtan uca test edilebilmeli (tahmin + forward-test kaydı).
+- FR-B5: Milli maç kalibrasyonu ayrı izlenmeli (kulüple karışmasın) — Model Performansı
+  sayfasında "milli" kırılımı.
+
+## Fonksiyonel Olmayan Gereksinimler
+- NFR-1: Gerçek para — milli maç eşik değişiklikleri backtest/forward-test ile doğrulanmadan
+  canlı agresifleştirilmemeli.
+- NFR-2: Lig etiketleme mevcut hacmi düşürmemeli (gizleme değil işaretleme).
+- NFR-3: Değişiklikler additive; mevcut kulüp tahmin akışını bozmamalı.
+
+## Açık Sorular (sonraki tasarım turunda netleşecek)
+- OQ-1: Lig güven eşikleri tam olarak ne? (örn. Yüksek = ROI>+10% & N≥30 BET)
+- OQ-2: Milli maçlar için brain eşiği nasıl ayarlanacak — ayrı tier mi, data_quality
+  flag istisnası mı? Önce backtest: milli maçlarda mevcut motor kaç BET/ne ROI verirdi
+  (eşik gevşetilse)? Bu ölçülmeden tuning yapılmamalı.
+- OQ-3: Etiket nerede saklanacak: yeni tablo mı, mevcut league_tiers mı, config mi?
+- OQ-4: Dünya Kupası grup maçlarında lineup geç gelir — probable_xi cezası milli maçta
+  nasıl ele alınacak?
+
+## Sonraki Adım
+1. (Önce) Lig güven etiketleme → /sc:design veya doğrudan uygulama (veri hazır).
+2. (Sonra) Milli maç backtest'i: eşik gevşetildiğinde milli maçlarda ROI ne? → ona göre tuning.

mds/national-team-strategy.md

@@ -0,0 +1,70 @@
+# Milli Takım / Dünya Kupası — Bahis Stratejisi (Veri-Temelli)
+
+> Kaynak: 2.300 maçlık milli backtest (multi_backtest_20260602, /tmp/bt_natl.csv).
+> Tüm rakamlar offline simülasyon (production'a dokunulmadan, aynı veride kural testi).
+> Tarih: 2026-06.
+
+## Temel Bulgular (kanıtlanmış)
+1. **Kalibrasyon İYİ** (MS ECE 1.6, OU15 2.2) — model olasılıkları milli maçta da doğru.
+   Sorun kalibrasyon değil EDGE. Yani piyasa oranları da keskin; avantaj sadece
+   belirli segmentlerde var.
+2. **Sadece MS market'inde edge var.** OU/BTTS/HT/DC/OE hepsi "bet-all" ROI
+   −12%..−21% — milli maçta gol/skor marketleri güvenilmez, KAPATILMALI.
+3. **MS'te edge oran bandına + rekabet türüne bağlı:**
+   - Favori (oran<3): zararlı (−10..−18%). Milli favoriler takılır (rotasyon/motivasyon).
+   - Denk-üstü (oran 4-7): ELEME/HAZIRLIK'ta kârlı, TURNUVA'da zararlı.
+4. **Rekabet türü kritik faktör** (DB'de feature YOK, lig adından türetilir):
+   HAZIRLIK / ELEME / TURNUVA çok farklı davranır.
+
+## Grid + Kararlılık Testi (overfit'e karşı)
+En iyi kombolar (N>=150, MS market):
+| kural | N | hit% | ROI |
+|---|---|---|---|
+| 4.0-7.0 sadece ELEME | 585 | 25% | +23.1% |
+| 3.5-6.0 HAZ+ELE | 1021 | 25% | +14.5% |
+| **4.0-7.0 HAZ+ELE (SEÇİLEN)** | **865** | **24%** | **+17.1%** |
+| 3.0-6.0 HAZ+ELE | 1381 | 25% | +10.1% |
+
+**Kararlılık (en güçlü kanıt):** "4-7 sadece ELEME" eski yarı +22.1% / yeni yarı +24.0%
+→ iki bağımsız zaman diliminde de pozitif = overfit DEĞİL, sahada tutar.
+
+## TURNUVA/FİNAL farkı (Dünya Kupası finalleri için kritik)
+Turnuva (Avrupa Şamp, Copa America, Uluslar Ligi, Gold Cup, Asya/Afrika Kupası):
+- 4-7 bandı turnuvada ZARARLI (−8.9%) — elemenin tersi.
+- Sadece underdog 5+ kârlı (+51% ama n=274, oynak, şans payı yüksek).
+- Sebep: büyük turnuva finallerinde favoriler tutarlı, sürpriz az.
+
+## SEÇİLEN STRATEJİ (kullanıcı kararı)
+**Milli-maç gate kuralı:**
+- Market: SADECE MS (diğer tüm marketler milli maçta kapalı)
+- Oran bandı: 4.0 ≤ odds < 7.0
+- Rekabet türü: SADECE Hazırlık + Eleme
+- TURNUVA/FİNAL: bahis ÖNERME (sadece analiz/olasılık göster). Underdog +51%
+  cazip ama oynak/az-örneklem → gerçek paraya bağlanmadı (kullanıcı kararı).
+Beklenen: +17% ROI, ~865 bahis/2300 maç. Mevcut gate +0.9% idi → ~19x iyileşme.
+
+## Mimari Notu (uygulama için)
+- Sorun model değil → ayrı ML modeli GEREKSİZ (1898 maç zaten overfit riski; karar verildi: kurma).
+- Çözüm = betting brain'de milli-maça özel GATE (eğitim-sonrası kural katmanı).
+- Rekabet türü lig adından türetilir: 'hazırlık'→HAZIRLIK, 'eleme/play-off'→ELEME,
+  diğer→TURNUVA. Milli lig tespiti: qualified_leagues.json'a eklenen 21 milli lig.
+- Kalıcı feature olarak rekabet türü eklenebilir (daha temiz) ama gate hardcode de yeter.
+
+## Durum: UYGULANDI + DOĞRULANDI (betting_brain v31f-national-regime).
+Kod:
+- utils/national_leagues.py — loader (data/national_leagues.json, 21 lig) + classify_competition
+- single_match_orchestrator.py — self.national_leagues yüklenir
+- orchestrator/market_board.py — match_info.is_national + competition_type; _is_national_match/_competition_type_for helpers
+- betting_brain.py _judge_row — national regime bloğu: is_national ise club mantığını override eder,
+  SADECE MS + 4.0-7.0 + (HAZIRLIK|ELEME) → BET (NATIONAL_BASE_SCORE 66, stake 0.5u, grade B),
+  diğer her şey REJECT. Hard-safety vetoları (low_reliability_hard, v25_v27_hard, htft_reversal)
+  national'da da geçerli. Rich analiz payload korunur.
+
+DOĞRULAMA (V2 backtest, yeni gate aktif, 1829 maç, /tmp/bt_natl_v2.csv):
+  BET=784 → TAMAMI MS, oran 4.00-6.99 (bant dışı 0 bahis), hit %23.7, ROI +16.0%, +125.7u.
+  Simülasyondaki +17% ile birebir. OU/BTTS/HT/turnuva artık 0 BET.
+
+NOT: ai-engine ~10:10'da restart oldu (compose) → national-gate + V31e recal + league_confidence
+  kodu CANLI API'de aktif. Ama bunlar docker cp ile deploy edildi; kalıcılık için repo commit +
+  image rebuild gerekir (yeni container build'inde kaybolur).
+## İlgili: 422 lig-gate düzeltmesi CANLIDA (qualified_leagues 48→69, milli ligler açıldı).

package-lock.json

@@ -27,6 +27,7 @@
         "@nestjs/terminus": "^11.0.0",
         "@nestjs/throttler": "^6.5.0",
         "@prisma/client": "^6.19.3",
+        "@resvg/resvg-js": "^2.6.2",
         "axios": "^1.13.6",
         "bcrypt": "^6.0.0",
         "bullmq": "^5.66.4",
@@ -49,6 +50,8 @@
         "prisma": "^6.19.3",
         "reflect-metadata": "^0.2.2",
         "rxjs": "^7.8.1",
+        "satori": "^0.26.0",
+        "satori-html": "^0.3.2",
         "twitter-api-v2": "^1.29.0",
         "zod": "^4.3.5"
       },
@@ -3912,12 +3915,243 @@
         "@redis/client": "^1.0.0"
       }
     },
+    "node_modules/@resvg/resvg-js": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js/-/resvg-js-2.6.2.tgz",
+      "integrity": "sha512-xBaJish5OeGmniDj9cW5PRa/PtmuVU3ziqrbr5xJj901ZDN4TosrVaNZpEiLZAxdfnhAe7uQ7QFWfjPe9d9K2Q==",
+      "license": "MPL-2.0",
+      "engines": {
+        "node": ">= 10"
+      },
+      "optionalDependencies": {
+        "@resvg/resvg-js-android-arm-eabi": "2.6.2",
+        "@resvg/resvg-js-android-arm64": "2.6.2",
+        "@resvg/resvg-js-darwin-arm64": "2.6.2",
+        "@resvg/resvg-js-darwin-x64": "2.6.2",
+        "@resvg/resvg-js-linux-arm-gnueabihf": "2.6.2",
+        "@resvg/resvg-js-linux-arm64-gnu": "2.6.2",
+        "@resvg/resvg-js-linux-arm64-musl": "2.6.2",
+        "@resvg/resvg-js-linux-x64-gnu": "2.6.2",
+        "@resvg/resvg-js-linux-x64-musl": "2.6.2",
+        "@resvg/resvg-js-win32-arm64-msvc": "2.6.2",
+        "@resvg/resvg-js-win32-ia32-msvc": "2.6.2",
+        "@resvg/resvg-js-win32-x64-msvc": "2.6.2"
+      }
+    },
+    "node_modules/@resvg/resvg-js-android-arm-eabi": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-android-arm-eabi/-/resvg-js-android-arm-eabi-2.6.2.tgz",
+      "integrity": "sha512-FrJibrAk6v29eabIPgcTUMPXiEz8ssrAk7TXxsiZzww9UTQ1Z5KAbFJs+Z0Ez+VZTYgnE5IQJqBcoSiMebtPHA==",
+      "cpu": [
+        "arm"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "android"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-android-arm64": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-android-arm64/-/resvg-js-android-arm64-2.6.2.tgz",
+      "integrity": "sha512-VcOKezEhm2VqzXpcIJoITuvUS/fcjIw5NA/w3tjzWyzmvoCdd+QXIqy3FBGulWdClvp4g+IfUemigrkLThSjAQ==",
+      "cpu": [
+        "arm64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "android"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-darwin-arm64": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-darwin-arm64/-/resvg-js-darwin-arm64-2.6.2.tgz",
+      "integrity": "sha512-nmok2LnAd6nLUKI16aEB9ydMC6Lidiiq2m1nEBDR1LaaP7FGs4AJ90qDraxX+CWlVuRlvNjyYJTNv8qFjtL9+A==",
+      "cpu": [
+        "arm64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "darwin"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-darwin-x64": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-darwin-x64/-/resvg-js-darwin-x64-2.6.2.tgz",
+      "integrity": "sha512-GInyZLjgWDfsVT6+SHxQVRwNzV0AuA1uqGsOAW+0th56J7Nh6bHHKXHBWzUrihxMetcFDmQMAX1tZ1fZDYSRsw==",
+      "cpu": [
+        "x64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "darwin"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-linux-arm-gnueabihf": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-linux-arm-gnueabihf/-/resvg-js-linux-arm-gnueabihf-2.6.2.tgz",
+      "integrity": "sha512-YIV3u/R9zJbpqTTNwTZM5/ocWetDKGsro0SWp70eGEM9eV2MerWyBRZnQIgzU3YBnSBQ1RcxRZvY/UxwESfZIw==",
+      "cpu": [
+        "arm"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-linux-arm64-gnu": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-linux-arm64-gnu/-/resvg-js-linux-arm64-gnu-2.6.2.tgz",
+      "integrity": "sha512-zc2BlJSim7YR4FZDQ8OUoJg5holYzdiYMeobb9pJuGDidGL9KZUv7SbiD4E8oZogtYY42UZEap7dqkkYuA91pg==",
+      "cpu": [
+        "arm64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-linux-arm64-musl": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-linux-arm64-musl/-/resvg-js-linux-arm64-musl-2.6.2.tgz",
+      "integrity": "sha512-3h3dLPWNgSsD4lQBJPb4f+kvdOSJHa5PjTYVsWHxLUzH4IFTJUAnmuWpw4KqyQ3NA5QCyhw4TWgxk3jRkQxEKg==",
+      "cpu": [
+        "arm64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-linux-x64-gnu": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-linux-x64-gnu/-/resvg-js-linux-x64-gnu-2.6.2.tgz",
+      "integrity": "sha512-IVUe+ckIerA7xMZ50duAZzwf1U7khQe2E0QpUxu5MBJNao5RqC0zwV/Zm965vw6D3gGFUl7j4m+oJjubBVoftw==",
+      "cpu": [
+        "x64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-linux-x64-musl": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-linux-x64-musl/-/resvg-js-linux-x64-musl-2.6.2.tgz",
+      "integrity": "sha512-UOf83vqTzoYQO9SZ0fPl2ZIFtNIz/Rr/y+7X8XRX1ZnBYsQ/tTb+cj9TE+KHOdmlTFBxhYzVkP2lRByCzqi4jQ==",
+      "cpu": [
+        "x64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-win32-arm64-msvc": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-win32-arm64-msvc/-/resvg-js-win32-arm64-msvc-2.6.2.tgz",
+      "integrity": "sha512-7C/RSgCa+7vqZ7qAbItfiaAWhyRSoD4l4BQAbVDqRRsRgY+S+hgS3in0Rxr7IorKUpGE69X48q6/nOAuTJQxeQ==",
+      "cpu": [
+        "arm64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "win32"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-win32-ia32-msvc": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-win32-ia32-msvc/-/resvg-js-win32-ia32-msvc-2.6.2.tgz",
+      "integrity": "sha512-har4aPAlvjnLcil40AC77YDIk6loMawuJwFINEM7n0pZviwMkMvjb2W5ZirsNOZY4aDbo5tLx0wNMREp5Brk+w==",
+      "cpu": [
+        "ia32"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "win32"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
+    "node_modules/@resvg/resvg-js-win32-x64-msvc": {
+      "version": "2.6.2",
+      "resolved": "https://registry.npmjs.org/@resvg/resvg-js-win32-x64-msvc/-/resvg-js-win32-x64-msvc-2.6.2.tgz",
+      "integrity": "sha512-ZXtYhtUr5SSaBrUDq7DiyjOFJqBVL/dOBN7N/qmi/pO0IgiWW/f/ue3nbvu9joWE5aAKDoIzy/CxsY0suwGosQ==",
+      "cpu": [
+        "x64"
+      ],
+      "license": "MPL-2.0",
+      "optional": true,
+      "os": [
+        "win32"
+      ],
+      "engines": {
+        "node": ">= 10"
+      }
+    },
     "node_modules/@scarf/scarf": {
       "version": "1.4.0",
       "resolved": "https://registry.npmjs.org/@scarf/scarf/-/scarf-1.4.0.tgz",
       "integrity": "sha512-xxeapPiUXdZAE3che6f3xogoJPeZgig6omHEy1rIY5WVsB3H2BHNnZH+gHG6x91SCWyQCzWGsuL2Hh3ClO5/qQ==",
       "hasInstallScript": true
     },
+    "node_modules/@shuding/opentype.js": {
+      "version": "1.4.0-beta.0",
+      "resolved": "https://registry.npmjs.org/@shuding/opentype.js/-/opentype.js-1.4.0-beta.0.tgz",
+      "integrity": "sha512-3NgmNyH3l/Hv6EvsWJbsvpcpUba6R8IREQ83nH83cyakCw7uM1arZKNfHwv1Wz6jgqrF/j4x5ELvR6PnK9nTcA==",
+      "license": "MIT",
+      "dependencies": {
+        "fflate": "^0.7.3",
+        "string.prototype.codepointat": "^0.2.1"
+      },
+      "bin": {
+        "ot": "bin/ot"
+      },
+      "engines": {
+        "node": ">= 8.0.0"
+      }
+    },
     "node_modules/@sinclair/typebox": {
       "version": "0.34.47",
       "resolved": "https://registry.npmjs.org/@sinclair/typebox/-/typebox-0.34.47.tgz",
@@ -6535,6 +6769,15 @@
         "node": ">=6"
       }
     },
+    "node_modules/camelize": {
+      "version": "1.0.1",
+      "resolved": "https://registry.npmjs.org/camelize/-/camelize-1.0.1.tgz",
+      "integrity": "sha512-dU+Tx2fsypxTgtLoE36npi3UqcjSSMNYfkqgmoEhtZrraP5VWq0K7FkWVTYa8eMPtnU/G2txVsfdCJTn9uzpuQ==",
+      "license": "MIT",
+      "funding": {
+        "url": "https://github.com/sponsors/ljharb"
+      }
+    },
     "node_modules/caniuse-lite": {
       "version": "1.0.30001762",
       "resolved": "https://registry.npmjs.org/caniuse-lite/-/caniuse-lite-1.0.30001762.tgz",
@@ -7086,6 +7329,36 @@
         "node": ">= 8"
       }
     },
+    "node_modules/css-background-parser": {
+      "version": "0.1.0",
+      "resolved": "https://registry.npmjs.org/css-background-parser/-/css-background-parser-0.1.0.tgz",
+      "integrity": "sha512-2EZLisiZQ+7m4wwur/qiYJRniHX4K5Tc9w93MT3AS0WS1u5kaZ4FKXlOTBhOjc+CgEgPiGY+fX1yWD8UwpEqUA==",
+      "license": "MIT"
+    },
+    "node_modules/css-box-shadow": {
+      "version": "1.0.0-3",
+      "resolved": "https://registry.npmjs.org/css-box-shadow/-/css-box-shadow-1.0.0-3.tgz",
+      "integrity": "sha512-9jaqR6e7Ohds+aWwmhe6wILJ99xYQbfmK9QQB9CcMjDbTxPZjwEmUQpU91OG05Xgm8BahT5fW+svbsQGjS/zPg==",
+      "license": "MIT"
+    },
+    "node_modules/css-color-keywords": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/css-color-keywords/-/css-color-keywords-1.0.0.tgz",
+      "integrity": "sha512-FyyrDHZKEjXDpNJYvVsV960FiqQyXc/LlYmsxl2BcdMb2WPx0OGRVgTg55rPSyLSNMqP52R9r8geSp7apN3Ofg==",
+      "license": "ISC",
+      "engines": {
+        "node": ">=4"
+      }
+    },
+    "node_modules/css-gradient-parser": {
+      "version": "0.0.17",
+      "resolved": "https://registry.npmjs.org/css-gradient-parser/-/css-gradient-parser-0.0.17.tgz",
+      "integrity": "sha512-w2Xy9UMMwlKtou0vlRnXvWglPAceXCTtcmVSo8ZBUvqCV5aXEFP/PC6d+I464810I9FT++UACwTD5511bmGPUg==",
+      "license": "MIT",
+      "engines": {
+        "node": ">=16"
+      }
+    },
     "node_modules/css-select": {
       "version": "5.2.2",
       "resolved": "https://registry.npmjs.org/css-select/-/css-select-5.2.2.tgz",
@@ -7102,6 +7375,17 @@
         "url": "https://github.com/sponsors/fb55"
       }
     },
+    "node_modules/css-to-react-native": {
+      "version": "3.2.0",
+      "resolved": "https://registry.npmjs.org/css-to-react-native/-/css-to-react-native-3.2.0.tgz",
+      "integrity": "sha512-e8RKaLXMOFii+02mOlqwjbD00KSEKqblnpO9e++1aXS1fPQOpS1YoqdVHBqPjHNoxeF2mimzVqawm2KCbEdtHQ==",
+      "license": "MIT",
+      "dependencies": {
+        "camelize": "^1.0.0",
+        "css-color-keywords": "^1.0.0",
+        "postcss-value-parser": "^4.0.2"
+      }
+    },
     "node_modules/css-what": {
       "version": "6.2.2",
       "resolved": "https://registry.npmjs.org/css-what/-/css-what-6.2.2.tgz",
@@ -7433,6 +7717,15 @@
       "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
       "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A=="
     },
+    "node_modules/emoji-regex-xs": {
+      "version": "2.0.1",
+      "resolved": "https://registry.npmjs.org/emoji-regex-xs/-/emoji-regex-xs-2.0.1.tgz",
+      "integrity": "sha512-1QFuh8l7LqUcKe24LsPUNzjrzJQ7pgRwp1QMcZ5MX6mFplk2zQ08NVCM84++1cveaUUYtcCYHmeFEuNg16sU4g==",
+      "license": "MIT",
+      "engines": {
+        "node": ">=10.0.0"
+      }
+    },
     "node_modules/empathic": {
       "version": "2.0.0",
       "resolved": "https://registry.npmjs.org/empathic/-/empathic-2.0.0.tgz",
@@ -8152,6 +8445,12 @@
         "node": "^12.20 || >= 14.13"
       }
     },
+    "node_modules/fflate": {
+      "version": "0.7.4",
+      "resolved": "https://registry.npmjs.org/fflate/-/fflate-0.7.4.tgz",
+      "integrity": "sha512-5u2V/CDW15QM1XbbgS+0DfPxVB+jUKhWEKuuFuHncbk3tEEqzmoXL+2KyOFuKGqOnmdIy0/davWF1CkuwtibCw==",
+      "license": "MIT"
+    },
     "node_modules/file-entry-cache": {
       "version": "8.0.0",
       "resolved": "https://registry.npmjs.org/file-entry-cache/-/file-entry-cache-8.0.0.tgz",
@@ -8794,6 +9093,18 @@
       "integrity": "sha512-7xgomUX6ADmcYzFik0HzAxh/73YlKR9bmFzf51CZwR+b6YtzU2m0u49hQCqV6SvlqIqsaxovfwdvbnsw3b/zpg==",
       "dev": true
     },
+    "node_modules/hex-rgb": {
+      "version": "4.3.0",
+      "resolved": "https://registry.npmjs.org/hex-rgb/-/hex-rgb-4.3.0.tgz",
+      "integrity": "sha512-Ox1pJVrDCyGHMG9CFg1tmrRUMRPRsAWYc/PinY0XzJU4K7y7vjNoLKIQ7BR5UJMCxNN8EM1MNDmHWA/B3aZUuw==",
+      "license": "MIT",
+      "engines": {
+        "node": ">=6"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/hookified": {
       "version": "1.15.0",
       "resolved": "https://registry.npmjs.org/hookified/-/hookified-1.15.0.tgz",
@@ -10097,6 +10408,25 @@
       "resolved": "https://registry.npmjs.org/libphonenumber-js/-/libphonenumber-js-1.12.33.tgz",
       "integrity": "sha512-r9kw4OA6oDO4dPXkOrXTkArQAafIKAU71hChInV4FxZ69dxCfbwQGDPzqR5/vea94wU705/3AZroEbSoeVWrQw=="
     },
+    "node_modules/linebreak": {
+      "version": "1.1.0",
+      "resolved": "https://registry.npmjs.org/linebreak/-/linebreak-1.1.0.tgz",
+      "integrity": "sha512-MHp03UImeVhB7XZtjd0E4n6+3xr5Dq/9xI/5FptGk5FrbDR3zagPa2DS6U8ks/3HjbKWG9Q1M2ufOzxV2qLYSQ==",
+      "license": "MIT",
+      "dependencies": {
+        "base64-js": "0.0.8",
+        "unicode-trie": "^2.0.0"
+      }
+    },
+    "node_modules/linebreak/node_modules/base64-js": {
+      "version": "0.0.8",
+      "resolved": "https://registry.npmjs.org/base64-js/-/base64-js-0.0.8.tgz",
+      "integrity": "sha512-3XSA2cR/h/73EzlXXdU6YNycmYI7+kicTxks4eJg2g39biHR84slg2+des+p7iHYhbRg/udIS4TD53WabcOUkw==",
+      "license": "MIT",
+      "engines": {
+        "node": ">= 0.4"
+      }
+    },
     "node_modules/lines-and-columns": {
       "version": "1.2.4",
       "resolved": "https://registry.npmjs.org/lines-and-columns/-/lines-and-columns-1.2.4.tgz",
@@ -11030,6 +11360,12 @@
       "resolved": "https://registry.npmjs.org/package-json-from-dist/-/package-json-from-dist-1.0.1.tgz",
       "integrity": "sha512-UEZIS3/by4OC8vL3P2dTXRETpebLI2NiI5vIrjaD/5UtrkFX/tNbwjTSRAGC/+7CAo2pIcBaRgWmcBBHcsaCIw=="
     },
+    "node_modules/pako": {
+      "version": "0.2.9",
+      "resolved": "https://registry.npmjs.org/pako/-/pako-0.2.9.tgz",
+      "integrity": "sha512-NUcwaKxUxWrZLpDG+z/xZaCgQITkA/Dv4V/T6bw7VON6l1Xz/VnrBqrYjZQ12TamKHzITTfOEIYUj48y2KXImA==",
+      "license": "MIT"
+    },
     "node_modules/parent-module": {
       "version": "1.0.1",
       "resolved": "https://registry.npmjs.org/parent-module/-/parent-module-1.0.1.tgz",
@@ -11042,6 +11378,16 @@
         "node": ">=6"
       }
     },
+    "node_modules/parse-css-color": {
+      "version": "0.2.1",
+      "resolved": "https://registry.npmjs.org/parse-css-color/-/parse-css-color-0.2.1.tgz",
+      "integrity": "sha512-bwS/GGIFV3b6KS4uwpzCFj4w297Yl3uqnSgIPsoQkx7GMLROXfMnWvxfNkL0oh8HVhZA4hvJoEoEIqonfJ3BWg==",
+      "license": "MIT",
+      "dependencies": {
+        "color-name": "^1.1.4",
+        "hex-rgb": "^4.1.0"
+      }
+    },
     "node_modules/parse-json": {
       "version": "5.2.0",
       "resolved": "https://registry.npmjs.org/parse-json/-/parse-json-5.2.0.tgz",
@@ -11438,6 +11784,12 @@
         "node": ">=4"
       }
     },
+    "node_modules/postcss-value-parser": {
+      "version": "4.2.0",
+      "resolved": "https://registry.npmjs.org/postcss-value-parser/-/postcss-value-parser-4.2.0.tgz",
+      "integrity": "sha512-1NNCs6uurfkVbeXG4S8JFT9t19m45ICnif8zWLd5oPSZ50QnwMfK+H3jv408d4jw/7Bttv5axS5IiHoLaVNHeQ==",
+      "license": "MIT"
+    },
     "node_modules/prebuild-install": {
       "version": "7.1.3",
       "resolved": "https://registry.npmjs.org/prebuild-install/-/prebuild-install-7.1.3.tgz",
@@ -12012,6 +12364,37 @@
       "resolved": "https://registry.npmjs.org/safer-buffer/-/safer-buffer-2.1.2.tgz",
       "integrity": "sha512-YZo3K82SD7Riyi0E1EQPojLz7kpepnSQI9IyPbHHg1XXXevb5dJI7tpyN2ADxGcQbHG7vcyRHk0cbwqcQriUtg=="
     },
+    "node_modules/satori": {
+      "version": "0.26.0",
+      "resolved": "https://registry.npmjs.org/satori/-/satori-0.26.0.tgz",
+      "integrity": "sha512-tkMFrfIs3l2mQ2JEcyW0ADTy3zGggFRFzi6Ef8YozQSFsFKEqaSO1Y8F9wJg4//PJGQauMalHGTUEkPrFwhVPA==",
+      "license": "MPL-2.0",
+      "dependencies": {
+        "@shuding/opentype.js": "1.4.0-beta.0",
+        "css-background-parser": "^0.1.0",
+        "css-box-shadow": "1.0.0-3",
+        "css-gradient-parser": "^0.0.17",
+        "css-to-react-native": "^3.0.0",
+        "emoji-regex-xs": "^2.0.1",
+        "escape-html": "^1.0.3",
+        "linebreak": "^1.1.0",
+        "parse-css-color": "^0.2.1",
+        "postcss-value-parser": "^4.2.0",
+        "yoga-layout": "^3.2.1"
+      },
+      "engines": {
+        "node": ">=16"
+      }
+    },
+    "node_modules/satori-html": {
+      "version": "0.3.2",
+      "resolved": "https://registry.npmjs.org/satori-html/-/satori-html-0.3.2.tgz",
+      "integrity": "sha512-wjTh14iqADFKDK80e51/98MplTGfxz2RmIzh0GqShlf4a67+BooLywF17TvJPD6phO0Hxm7Mf1N5LtRYvdkYRA==",
+      "license": "MIT",
+      "dependencies": {
+        "ultrahtml": "^1.2.0"
+      }
+    },
     "node_modules/schema-utils": {
       "version": "3.3.0",
       "resolved": "https://registry.npmjs.org/schema-utils/-/schema-utils-3.3.0.tgz",
@@ -12488,6 +12871,12 @@
         "node": ">=8"
       }
     },
+    "node_modules/string.prototype.codepointat": {
+      "version": "0.2.1",
+      "resolved": "https://registry.npmjs.org/string.prototype.codepointat/-/string.prototype.codepointat-0.2.1.tgz",
+      "integrity": "sha512-2cBVCj6I4IOvEnjgO/hWqXjqBGsY+zwPmHl12Srk9IXSZ56Jwwmy+66XO5Iut/oQVR7t5ihYdLB0GMa4alEUcg==",
+      "license": "MIT"
+    },
     "node_modules/strip-ansi": {
       "version": "6.0.1",
       "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
@@ -12858,6 +13247,12 @@
         "real-require": "^0.2.0"
       }
     },
+    "node_modules/tiny-inflate": {
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/tiny-inflate/-/tiny-inflate-1.0.3.tgz",
+      "integrity": "sha512-pkY1fj1cKHb2seWDy0B16HeWyczlJA9/WW3u3c4z/NiWDsO3DOU5D7nhTLE9CF0yXv/QZFY7sEJmj24dK+Rrqw==",
+      "license": "MIT"
+    },
     "node_modules/tinyexec": {
       "version": "1.1.2",
       "resolved": "https://registry.npmjs.org/tinyexec/-/tinyexec-1.1.2.tgz",
@@ -13260,6 +13655,12 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
+    "node_modules/ultrahtml": {
+      "version": "1.6.0",
+      "resolved": "https://registry.npmjs.org/ultrahtml/-/ultrahtml-1.6.0.tgz",
+      "integrity": "sha512-R9fBn90VTJrqqLDwyMph+HGne8eqY1iPfYhPzZrvKpIfwkWZbcYlfpsb8B9dTvBfpy1/hqAD7Wi8EKfP9e8zdw==",
+      "license": "MIT"
+    },
     "node_modules/undici": {
       "version": "7.18.2",
       "resolved": "https://registry.npmjs.org/undici/-/undici-7.18.2.tgz",
@@ -13274,6 +13675,16 @@
       "resolved": "https://registry.npmjs.org/undici-types/-/undici-types-6.21.0.tgz",
       "integrity": "sha512-iwDZqg0QAGrg9Rav5H4n0M64c3mkR59cJ6wQp+7C4nI0gsmExaedaYLNO44eT4AtBBwjbTiGPMlt2Md0T9H9JQ=="
     },
+    "node_modules/unicode-trie": {
+      "version": "2.0.0",
+      "resolved": "https://registry.npmjs.org/unicode-trie/-/unicode-trie-2.0.0.tgz",
+      "integrity": "sha512-x7bc76x0bm4prf1VLg79uhAzKw8DVboClSN5VxJuQ+LKDOVEW9CdH+VY7SP+vX7xCYQqzzgQpFqz15zeLvAtZQ==",
+      "license": "MIT",
+      "dependencies": {
+        "pako": "^0.2.5",
+        "tiny-inflate": "^1.0.0"
+      }
+    },
     "node_modules/universalify": {
       "version": "2.0.1",
       "resolved": "https://registry.npmjs.org/universalify/-/universalify-2.0.1.tgz",
@@ -13858,6 +14269,12 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
+    "node_modules/yoga-layout": {
+      "version": "3.2.1",
+      "resolved": "https://registry.npmjs.org/yoga-layout/-/yoga-layout-3.2.1.tgz",
+      "integrity": "sha512-0LPOt3AxKqMdFBZA3HBAt/t/8vIKq7VaQYbuA8WxCgung+p9TVyKRYdpvCb80HcdTN2NkbIKbhNwKUfm3tQywQ==",
+      "license": "MIT"
+    },
     "node_modules/zod": {
       "version": "4.3.5",
       "resolved": "https://registry.npmjs.org/zod/-/zod-4.3.5.tgz",

package.json

@@ -60,6 +60,7 @@
     "@nestjs/terminus": "^11.0.0",
     "@nestjs/throttler": "^6.5.0",
     "@prisma/client": "^6.19.3",
+    "@resvg/resvg-js": "^2.6.2",
     "axios": "^1.13.6",
     "bcrypt": "^6.0.0",
     "bullmq": "^5.66.4",
@@ -82,6 +83,8 @@
     "prisma": "^6.19.3",
     "reflect-metadata": "^0.2.2",
     "rxjs": "^7.8.1",
+    "satori": "^0.26.0",
+    "satori-html": "^0.3.2",
     "twitter-api-v2": "^1.29.0",
     "zod": "^4.3.5"
   },
@@ -134,4 +137,4 @@
   "prisma": {
     "seed": "ts-node prisma/seed.ts"
   }
-}
+}

parse_errors.py

@@ -1,17 +0,0 @@
-import json
-
-targets = [
-    "bet_recommender.py", "score_calculator.py", "db.py", "upset_engine_v2.py",
-    "v20_ensemble.py", "v27_predictor.py", "betting_brain.py",
-    "single_match_orchestrator.py", "v26_shadow_engine.py"
-]
-
-d = json.load(open("pyright_main_errors.json", encoding="utf-16"))
-for diag in d["generalDiagnostics"]:
-    if diag["severity"] == "error":
-        fname = diag["file"]
-        if any(t in fname for t in targets):
-            # Print safely encoding to ascii to avoid charmap errors
-            safe_fname = fname.split('ai-engine')[1].encode('ascii', 'ignore').decode()
-            safe_msg = diag["message"].encode('ascii', 'ignore').decode()
-            print(f"{safe_fname} L{diag['range']['start']['line']+1}: {safe_msg}")

predict-single.ts

@@ -1,153 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-import axios from 'axios';
-
-dotenv.config();
-
-// BigInt serialization fix
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-const matchId = '30gnuehy43on5orc9n3sh8pw4'; // Valencia vs Celta Vigo - HT/FT test
-
-async function getPrediction() {
-  console.log('🔮 VQWEN v3 PREDICTION');
-  console.log('='.repeat(80));
-
-  // Fetch match from database
-  const match = await prisma.liveMatch.findUnique({
-    where: { id: matchId },
-    include: {
-      homeTeam: true,
-      awayTeam: true,
-      league: true,
-    },
-  });
-
-  if (!match) {
-    console.log(`❌ Match not found: ${matchId}`);
-    await prisma.$disconnect();
-    return;
-  }
-
-  console.log(`\n📊 ${match.homeTeam?.name} vs ${match.awayTeam?.name}`);
-  console.log(`🏆 League: ${match.league?.name}`);
-  console.log(`📅 Match Time: ${new Date(Number(match.mstUtc)).toISOString()}`);
-  console.log(`📍 Status: ${match.state} / ${match.substate}`);
-
-  // Check data availability
-  console.log(`\n📦 DATA CHECK:`);
-  console.log(`   Odds: ${match.odds ? '✅' : '❌'}`);
-  console.log(`   Lineups: ${match.lineups ? '✅' : '❌'}`);
-  console.log(`   Sidelined: ${match.sidelined ? '✅' : '❌'}`);
-  console.log(`   Referee: ${match.refereeName || 'N/A'}`);
-
-  // Send prediction request
-  const aiEngineUrl = 'http://localhost:8007';
-  const predictionUrl = `${aiEngineUrl}/v20plus/analyze/${matchId}`;
-
-  console.log(`\n🤖 Sending to AI Engine...`);
-
-  const startTime = Date.now();
-  const response = await axios.post(predictionUrl, {}, {
-    timeout: 120000,
-  });
-  const elapsed = ((Date.now() - startTime) / 1000).toFixed(2);
-
-  console.log(`✅ Prediction received in ${elapsed}s\n`);
-
-  const pkg = response.data;
-
-  // Print full JSON
-  console.log('='.repeat(80));
-  console.log('📊 FULL PREDICTION JSON:');
-  console.log('='.repeat(80));
-  console.log(JSON.stringify(pkg, null, 2));
-
-  // Summary
-  console.log(`\n${'='.repeat(80)}`);
-  console.log('🎯 PREDICTION SUMMARY:');
-  console.log('='.repeat(80));
-
-  const dq = pkg.data_quality;
-  console.log(`\n📦 Data Quality: ${dq.label} (${dq.score})`);
-  console.log(`   Home lineup: ${dq.home_lineup_count}`);
-  console.log(`   Away lineup: ${dq.away_lineup_count}`);
-  console.log(`   Source: ${dq.lineup_source}`);
-
-  const eb = pkg.engine_breakdown;
-  console.log(`\n📈 Engine Signals:`);
-  console.log(`   Team: ${eb.team}%`);
-  console.log(`   Player: ${eb.player}%`);
-  console.log(`   Odds: ${eb.odds}%`);
-  console.log(`   Referee: ${eb.referee}%`);
-
-  const mp = pkg.main_pick;
-  console.log(`\n🥇 Main Pick:`);
-  console.log(`   Market: ${mp.market}`);
-  console.log(`   Pick: ${mp.pick}`);
-  console.log(`   Confidence: ${mp.confidence}%`);
-  console.log(`   Odds: ${mp.odds}`);
-  console.log(`   Edge: ${(mp.edge * 100).toFixed(2)}%`);
-  console.log(`   Grade: ${mp.bet_grade}`);
-  console.log(`   Playable: ${mp.playable}`);
-  console.log(`   Stake: ${mp.stake_units} units`);
-
-  if (pkg.value_pick) {
-    const vp = pkg.value_pick;
-    console.log(`\n💎 Value Pick:`);
-    console.log(`   Market: ${vp.market}`);
-    console.log(`   Pick: ${vp.pick}`);
-    console.log(`   Confidence: ${vp.confidence}%`);
-    console.log(`   Odds: ${vp.odds}`);
-    console.log(`   Edge: ${(vp.edge * 100).toFixed(2)}%`);
-  }
-
-  const sp = pkg.score_prediction;
-  console.log(`\n⚽ Score Prediction:`);
-  console.log(`   FT: ${sp.ft}`);
-  console.log(`   HT: ${sp.ht}`);
-  console.log(`   xG: ${sp.xg_home} - ${sp.xg_away} (Total: ${sp.xg_total})`);
-
-  console.log(`\n🎲 Top 5 Scores:`);
-  pkg.scenario_top5.forEach((s: any, i: number) => {
-    console.log(`   ${i + 1}. ${s.score} (${s.prob}%)`);
-  });
-
-  const risk = pkg.risk;
-  console.log(`\n⚠️ Risk: ${risk.level} (${risk.score}/10)`);
-  if (risk.warnings?.length > 0) {
-    console.log(`   Warnings: ${risk.warnings.join(', ')}`);
-  }
-
-  console.log(`\n💬 Reasoning:`);
-  pkg.reasoning_factors.forEach((f: string) => console.log(`   - ${f}`));
-
-  if (pkg.ai_commentary) {
-    console.log(`\n💬 AI Commentary:`);
-    console.log(`   ${pkg.ai_commentary}`);
-  }
-
-  // HT/FT specific check
-  console.log(`\n🔍 HT/FT CHECK:`);
-  const htft = pkg.market_board?.HTFT;
-  if (htft && htft.probs && Object.keys(htft.probs).length > 0) {
-    console.log(`   ✅ HT/FT PROBS PRESENT:`);
-    Object.entries(htft.probs).forEach(([key, val]) => {
-      console.log(`      ${key}: ${(val as number * 100).toFixed(2)}%`);
-    });
-
-    // Find best HT/FT
-    const best = Object.entries(htft.probs).reduce((a, b) => (b[1] as number) > (a[1] as number) ? b : a);
-    console.log(`\n   🎯 BEST HT/FT: ${best[0]} (${(best[1] as number * 100).toFixed(2)}%)`);
-  } else {
-    console.log(`   ❌ HT/FT PROBS EMPTY`);
-  }
-
-  await prisma.$disconnect();
-}
-
-getPrediction().catch(console.error);

qualified_leagues.json

@@ -46,5 +46,27 @@
   "722fdbecxzcq9788l6jqclzlw",
   "8ey0ww2zsosdmwr8ehsorh6t7",
   "32n2r9bl6x90psj0wa7bfs6vq",
-  "59tpnfrwnvhnhzmnvfyug68hj"
+  "59tpnfrwnvhnhzmnvfyug68hj",
+  "cesdwwnxbc5fmajgroc0hqzy2",
+  "3aa4mumjl6zyetg6o9hwd5hhx",
+  "40yjcbx2sq6oq736iqqqczwt1",
+  "39q1hq42hxjfylxb7xpe9bvf9",
+  "cu0rmpyff5692eo06ltddjo8a",
+  "ax1yf4nlzqpcji4j8epdgx3zl",
+  "1gxlzw2ezkyeykhcaa5x8ozkk",
+  "gfskxsdituog2kqp9yiu7bzi",
+  "595nsvo7ykvoe690b1e4u5n56",
+  "68zplepppndhl8bfdvgy9vgu1",
+  "3a0j0giz3c3ajw9h59evv7lqt",
+  "emy1ibc8fu2l0fukh4vlu5xl5",
+  "2db0aw1duj2my9l5iey5gm6nq",
+  "cc5tzz23tryrfqbm2pbv0jill",
+  "8tddm56zbasf57jkkay4kbf11",
+  "2r1hqz453bn9ljzt53kdr2lwb",
+  "93i7thp7zi0ympyt6l8aa1r2i",
+  "45db8orh1qttbsqq9hqapmbit",
+  "ude9t6yj60lebbn356qzg4k4",
+  "9qzn8cs96sgtqmesa9gpfti23",
+  "ad8y7vdjhinfqv4wo8rod6dck",
+  "70excpe1synn9kadnbppahdn7"
 ]

query-3-matches.ts

@@ -1,74 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-
-dotenv.config();
-
-// BigInt serialization fix
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-const matchIds = [
-  '7cnm7h7qbsq2bbaxngusojh90',
-  '7lmrfu2k1e2uxprxfxgaevcb8',
-  '3ko3otchy41d28rzxfpvl3d3o'
-];
-
-async function getMatches() {
-  for (const matchId of matchIds) {
-    try {
-      const match = await prisma.liveMatch.findUnique({
-        where: { id: matchId },
-        include: {
-          homeTeam: true,
-          awayTeam: true,
-          league: true,
-        },
-      });
-
-      if (!match) {
-        console.log(`\n❌ Maç bulunamadı: ${matchId}`);
-        continue;
-      }
-
-      console.log(`\n${'='.repeat(80)}`);
-      console.log(`📊 MAÇ: ${match.homeTeam?.name} vs ${match.awayTeam?.name}`);
-      console.log('='.repeat(80));
-      console.log(`ID: ${match.id}`);
-      console.log(`Lig: ${match.league?.name}`);
-      console.log(`Durum: ${match.state} / ${match.substate}`);
-      console.log(`Maç Zamanı (MS): ${match.mstUtc?.toString()}`);
-      console.log(`Hakem: ${match.refereeName || 'Bilinmiyor'}`);
-      console.log(`İlk 11 Var: ${match.lineups ? '✅' : '❌'}`);
-      console.log(`Sakat/Cezalı: ${match.sidelined ? '✅ Var' : '❌ Yok'}`);
-      
-      // Lineups summary
-      if (match.lineups) {
-        const lineups = match.lineups as any;
-        if (lineups.home && lineups.home.xi) {
-          console.log(`\n🏠 EV SAHİBİ İLK 11 (${match.homeTeam?.name}):`);
-          lineups.home.xi.forEach((p: any) => {
-            console.log(`   ${p.matchName} (${p.shirtNumber}) - ${p.position}`);
-          });
-        }
-        if (lineups.away && lineups.away.xi) {
-          console.log(`\n✈️ DEPLASMAN İLK 11 (${match.awayTeam?.name}):`);
-          lineups.away.xi.forEach((p: any) => {
-            console.log(`   ${p.matchName} (${p.shirtNumber}) - ${p.position}`);
-          });
-        }
-      }
-      
-      console.log('\n');
-      
-    } catch (error) {
-      console.error(`❌ Hata (${matchId}):`, error);
-    }
-  }
-  
-  await prisma.$disconnect();
-}
-
-getMatches();

query-match.ts

@@ -1,51 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-
-dotenv.config();
-
-const prisma = new PrismaClient();
-
-// BigInt serialization fix
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-async function getMatch() {
-  try {
-    const match = await prisma.liveMatch.findUnique({
-      where: { id: '3kemwubzpmga0nwhtc0o0vgno' },
-      include: {
-        homeTeam: true,
-        awayTeam: true,
-        league: true,
-      },
-    });
-
-    if (!match) {
-      console.log('❌ Maç bulunamadı!');
-      return;
-    }
-
-    console.log('✅ Maç bulundu:');
-    console.log(JSON.stringify(match, null, 2));
-
-    // Maç bilgilerini özetle
-    console.log('\n📊 MAÇ ÖZETİ:');
-    console.log('ID:', match.id);
-    console.log('Slug:', match.matchSlug);
-    console.log('Ev sahibi:', match.homeTeam?.name);
-    console.log('Deplasman:', match.awayTeam?.name);
-    console.log('Lig:', match.league?.name);
-    console.log('Durum:', match.status);
-    console.log('Spor:', match.sport);
-    console.log('Maç Zamanı (MS):', match.mstUtc?.toString());
-    console.log('Skor:', match.scoreHome, '-', match.scoreAway);
-
-  } catch (error) {
-    console.error('❌ Hata:', error);
-  } finally {
-    await prisma.$disconnect();
-  }
-}
-
-getMatch();

query-matches-detailed.ts

@@ -1,100 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-
-dotenv.config();
-
-// BigInt serialization fix
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-const matchIds = [
-  '7cnm7h7qbsq2bbaxngusojh90',
-  '7lmrfu2k1e2uxprxfxgaevcb8',
-  '3ko3otchy41d28rzxfpvl3d3o'
-];
-
-async function getMatches() {
-  for (const matchId of matchIds) {
-    try {
-      const match = await prisma.liveMatch.findUnique({
-        where: { id: matchId },
-        include: {
-          homeTeam: true,
-          awayTeam: true,
-          league: true,
-        },
-      });
-
-      if (!match) {
-        console.log(`\n❌ Maç bulunamadı: ${matchId}`);
-        continue;
-      }
-
-      console.log(`\n${'='.repeat(80)}`);
-      console.log(`🏟️ ${match.homeTeam?.name} vs ${match.awayTeam?.name}`);
-      console.log(`📍 Lig: ${match.league?.name}`);
-      console.log(`📅 Maç Zamanı: ${new Date(Number(match.mstUtc)).toLocaleString('tr-TR')}`);
-      console.log(`👨‍⚖️ Hakem: ${match.refereeName || 'Bilinmiyor'}`);
-      console.log('='.repeat(80));
-      
-      // Lineups
-      if (match.lineups) {
-        const lineups = match.lineups as any;
-        
-        // Home team
-        if (lineups.home && lineups.home.xi) {
-          console.log(`\n🏠 EV SAHİBİ İLK 11 (${match.homeTeam?.name}):`);
-          console.log('-'.repeat(80));
-          const goalscorers = lineups.home.xi.filter((p: any) => p.events?.some((e: any) => e.name === 'goal'));
-          const cards = lineups.home.xi.filter((p: any) => p.events?.some((e: any) => e.name === 'yellow-card' || e.name === 'red-card'));
-          const subs = lineups.home.xi.filter((p: any) => p.events?.some((e: any) => e.name === 'sub-off'));
-          
-          lineups.home.xi.forEach((p: any) => {
-            const hasGoal = p.events?.some((e: any) => e.name === 'goal');
-            const hasCard = p.events?.some((e: any) => e.name === 'yellow-card' || e.name === 'red-card');
-            const marker = hasGoal ? ' ⚽' : hasCard ? ' 🟨' : '';
-            console.log(`   ${p.matchName} (${p.shirtNumber}) - ${p.position}${marker}`);
-          });
-          
-          if (goalscorers.length > 0) {
-            console.log(`   ⚽ Gol Edenler: ${goalscorers.map((p: any) => p.matchName).join(', ')}`);
-          }
-        }
-        
-        // Away team
-        if (lineups.away && lineups.away.xi) {
-          console.log(`\n✈️ DEPLASMAN İLK 11 (${match.awayTeam?.name}):`);
-          console.log('-'.repeat(80));
-          lineups.away.xi.forEach((p: any) => {
-            const hasGoal = p.events?.some((e: any) => e.name === 'goal');
-            const hasCard = p.events?.some((e: any) => e.name === 'yellow-card' || e.name === 'red-card');
-            const marker = hasGoal ? ' ⚽' : hasCard ? ' 🟨' : '';
-            console.log(`   ${p.matchName} (${p.shirtNumber}) - ${p.position}${marker}`);
-          });
-        }
-        
-        // Sidelined
-        if (match.sidelined) {
-          const sidelined = match.sidelined as any;
-          const homeSidelined = sidelined.homeTeam?.totalSidelined || 0;
-          const awaySidelined = sidelined.awayTeam?.totalSidelined || 0;
-          console.log(`\n🏥 Sakat/Cezalı:`);
-          console.log(`   Ev Sahibi: ${homeSidelined} oyuncu`);
-          console.log(`   Deplasman: ${awaySidelined} oyuncu`);
-        }
-      }
-      
-      console.log('\n');
-      
-    } catch (error) {
-      console.error(`❌ Hata (${matchId}):`, error);
-    }
-  }
-  
-  await prisma.$disconnect();
-}
-
-getMatches();

query-matches-simple.ts

@@ -1,63 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-
-dotenv.config();
-
-// BigInt serialization fix
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-async function getMatches() {
-  const matches = await prisma.liveMatch.findMany({
-    where: {
-      id: {
-        in: [
-          '7cnm7h7qbsq2bbaxngusojh90',
-          '7lmrfu2k1e2uxprxfxgaevcb8',
-          '3ko3otchy41d28rzxfpvl3d3o'
-        ]
-      }
-    },
-    include: {
-      homeTeam: true,
-      awayTeam: true,
-      league: true,
-    },
-  });
-
-  matches.forEach((match, idx) => {
-    console.log(`\n${'='.repeat(80)}`);
-    console.log(`MAÇ ${idx + 1}: ${match.homeTeam?.name} vs ${match.awayTeam?.name}`);
-    console.log('='.repeat(80));
-    console.log(`ID: ${match.id}`);
-    console.log(`Lig: ${match.league?.name} (${match.league?.countryId})`);
-    console.log(`Durum: ${match.state} / ${match.substate}`);
-    console.log(`Skor: ${match.scoreHome ?? '?'} - ${match.scoreAway ?? '?'}`);
-    console.log(`Hakem: ${match.refereeName || 'Bilinmiyor'}`);
-    console.log(`Lineups Tip: ${typeof match.lineups} | ${match.lineups ? 'VAR' : 'YOK'}`);
-    
-    if (match.lineups) {
-      const lineups = match.lineups as any;
-      console.log(`Lineups Keys: ${Object.keys(lineups).join(', ')}`);
-      
-      // Check structure
-      if (lineups.home) {
-        const homeXi = lineups.home.xi || lineups.home.stats || [];
-        console.log(`Ev Sahibi İlk 11: ${Array.isArray(homeXi) ? homeXi.length : 'N/A'} oyuncu`);
-      }
-      if (lineups.away) {
-        const awayXi = lineups.away.xi || lineups.away.stats || [];
-        console.log(`Deplasman İlk 11: ${Array.isArray(awayXi) ? awayXi.length : 'N/A'} oyuncu`);
-      }
-    }
-    
-    console.log('');
-  });
-
-  await prisma.$disconnect();
-}
-
-getMatches();

run-prediction.py

@@ -1,170 +0,0 @@
-"""
-VQWEN v3 Model - Manual Prediction Script
-Match ID: 558o1fq1vbfsi3m5gm4ekpyc4
-Match: Kaiserslautern vs F. Düsseldorf
-League: 2. Bundesliga
-"""
-
-import requests
-import json
-from datetime import datetime
-
-# AI Engine base URL
-AI_ENGINE_URL = "http://127.0.0.1:8000"
-MATCH_ID = "558o1fq1vbfsi3m5gm4ekpyc4"
-
-def check_engine_health():
-    """Check if AI Engine is running"""
-    try:
-        response = requests.get(f"{AI_ENGINE_URL}/health", timeout=5)
-        return response.status_code == 200
-    except:
-        return False
-
-def run_prediction():
-    """Run VQWEN v3 prediction for the match"""
-    
-    print("=" * 80)
-    print("🤖 VQWEN v3 MODEL - MANUEL TAHMİN SİSTEMİ")
-    print("=" * 80)
-    print(f"\n📊 Maç Bilgileri:")
-    print(f"   ID: {MATCH_ID}")
-    print(f"   Ev Sahibi: Kaiserslautern")
-    print(f"   Deplasman: F. Düsseldorf")
-    print(f"   Lig: 2. Bundesliga")
-    print(f"   Maç Zamanı: 2026-04-01 (MS: 1775300400000)")
-    print(f"   Hakem: D. Schlager")
-    print()
-    
-    # Check engine health
-    print("🔍 AI Engine kontrol ediliyor...")
-    if not check_engine_health():
-        print("❌ AI Engine (Python FastAPI) çalışmıyor!")
-        print()
-        print("ℹ️  Lütfen AI Engine'i başlatın:")
-        print("   cd ai-engine")
-        print("   uvicorn main:app --host 0.0.0.0 --port 8000 --reload")
-        print()
-        print("📋 Alternatif olarak, maç verilerini hazırlayabilirim:")
-        print()
-        
-        # Prepare match data for analysis
-        match_data = {
-            "match_id": MATCH_ID,
-            "home_team": "Kaiserslautern",
-            "away_team": "F. Düsseldorf",
-            "league": "2. Bundesliga",
-            "match_date_ms": "1775300400000",
-            "referee": "D. Schlager",
-            "odds": {
-                "MS_1": 2.13,
-                "MS_X": 3.23,
-                "MS_2": 2.34,
-                "Alt_2.5": 2.09,
-                "Ust_2.5": 1.38,
-                "KG_Var": 1.32,
-                "KG_Yok": 2.25
-            },
-            "lineups_available": True,
-            "sidelined_count": 0
-        }
-        
-        print("✅ Maç verileri hazırlandı:")
-        print(json.dumps(match_data, indent=2, ensure_ascii=False))
-        print()
-        print("⚠️  Tahmin almak için AI Engine'in çalışması gerekiyor.")
-        print()
-        return
-    
-    # If engine is running, call the analysis endpoint
-    print("✅ AI Engine çalışıyor!")
-    print()
-    print("🎯 Tahmin yapılıyor...")
-    
-    try:
-        response = requests.post(
-            f"{AI_ENGINE_URL}/v20plus/analyze/{MATCH_ID}",
-            json={},
-            timeout=60
-        )
-        
-        if response.status_code == 200:
-            result = response.json()
-            
-            print("\n" + "=" * 80)
-            print("📊 TAHMİN SONUÇLARI")
-            print("=" * 80)
-            
-            # Main Pick
-            if 'main_pick' in result:
-                main = result['main_pick']
-                print(f"\n🎯 ANA TAHMİN:")
-                print(f"   Market: {main.get('market', 'N/A')}")
-                print(f"   Tahmin: {main.get('pick', 'N/A')}")
-                print(f"   Oran: {main.get('odds', 'N/A')}")
-                print(f"   Güven: {main.get('confidence', 0):.1f}%")
-                print(f"   Olasılık: {main.get('probability', 0):.1f}%")
-                print(f"   Bahis Derecesi: {main.get('bet_grade', 'N/A')}")
-            
-            # Value Pick
-            if 'value_pick' in result:
-                value = result['value_pick']
-                print(f"\n💎 DEĞER TAHMİNİ:")
-                print(f"   Market: {value.get('market', 'N/A')}")
-                print(f"   Tahmin: {value.get('pick', 'N/A')}")
-                print(f"   Oran: {value.get('odds', 'N/A')}")
-                print(f"   Güven: {value.get('confidence', 0):.1f}%")
-                print(f"   Edge: {value.get('edge', 0):.2f}")
-            
-            # Score Prediction
-            if 'score_prediction' in result:
-                score = result['score_prediction']
-                print(f"\n⚽ SKOR TAHMİNİ:")
-                print(f"   İlk Yarı: {score.get('ht', 'N/A')}")
-                print(f"   Maç Sonu: {score.get('ft', 'N/A')}")
-                print(f"   xG (Ev): {score.get('xg_home', 0):.2f}")
-                print(f"   xG (Dep): {score.get('xg_away', 0):.2f}")
-                print(f"   Toplam xG: {score.get('xg_total', 0):.2f}")
-            
-            # Bet Summary
-            if 'bet_summary' in result:
-                print(f"\n📋 TÜM TAHMİNLER:")
-                for bet in result['bet_summary']:
-                    print(f"   • {bet.get('market', 'N/A')}: {bet.get('pick', 'N/A')} "
-                          f"(Güven: {bet.get('calibrated_confidence', 0):.1f}%, "
-                          f"Derece: {bet.get('bet_grade', 'N/A')})")
-            
-            # AI Commentary
-            if 'ai_commentary' in result:
-                print(f"\n💬 AI YORUMU:")
-                print(f"   {result['ai_commentary']}")
-            
-            # Risk Assessment
-            if 'risk' in result:
-                risk = result['risk']
-                print(f"\n⚠️  RİSK DEĞERLENDİRMESİ:")
-                print(f"   Seviye: {risk.get('level', 'N/A')}")
-                print(f"   Skor: {risk.get('score', 0):.1f}")
-                if risk.get('warnings'):
-                    print(f"   Uyarılar: {', '.join(risk['warnings'][:3])}")
-            
-            # Data Quality
-            if 'data_quality' in result:
-                quality = result['data_quality']
-                print(f"\n📊 VERİ KALİTESİ:")
-                print(f"   Seviye: {quality.get('label', 'N/A')}")
-                print(f"   Skor: {quality.get('score', 0):.1f}")
-            
-            print("\n" + "=" * 80)
-            
-        else:
-            print(f"❌ Hata: HTTP {response.status_code}")
-            print(f"   {response.text}")
-            
-    except requests.exceptions.Timeout:
-        print("❌ Zaman aşımı! AI Engine yanıt vermiyor.")
-    except Exception as e:
-        print(f"❌ Hata: {str(e)}")
-
-if __name__ == "__main__":
-    run_prediction()

run-predictions.ts

@@ -1,153 +0,0 @@
-import { PrismaClient } from '@prisma/client';
-import * as dotenv from 'dotenv';
-import axios from 'axios';
-
-dotenv.config();
-
-// BigInt serialization fix
-(BigInt.prototype as any).toJSON = function () {
-  return this.toString();
-};
-
-const prisma = new PrismaClient();
-
-const matchIds = [
-  '7cnm7h7qbsq2bbaxngusojh90',  // Club Brugge vs Anderlecht - TESTED ✅
-  '7lmrfu2k1e2uxprxfxgaevcb8',  // Castellon vs Granada
-  '3ko3otchy41d28rzxfpvl3d3o'  // SV Ried vs Altach
-];
-
-async function getPrediction(matchId: string) {
-  try {
-    console.log(`\n${'='.repeat(80)}`);
-    console.log(`🔮 PREDICTION REQUEST: ${matchId}`);
-    console.log('='.repeat(80));
-
-    // Fetch match from database
-    const match = await prisma.liveMatch.findUnique({
-      where: { id: matchId },
-      include: {
-        homeTeam: true,
-        awayTeam: true,
-        league: true,
-      },
-    });
-
-    if (!match) {
-      console.log(`❌ Match not found: ${matchId}`);
-      return null;
-    }
-
-    console.log(`📊 ${match.homeTeam?.name} vs ${match.awayTeam?.name}`);
-    console.log(`🏆 League: ${match.league?.name}`);
-    console.log(`📅 Match Time: ${new Date(Number(match.mstUtc)).toISOString()}`);
-
-    // Send prediction request to AI Engine
-    const aiEngineUrl = 'http://localhost:8007';
-    const predictionUrl = `${aiEngineUrl}/v20plus/analyze/${matchId}`;
-
-    console.log(`\n🤖 Sending to AI Engine: ${predictionUrl}`);
-
-    const startTime = Date.now();
-    const response = await axios.post(predictionUrl, {}, {
-      timeout: 120000, // 2 minutes timeout
-    });
-
-    const elapsed = ((Date.now() - startTime) / 1000).toFixed(2);
-
-    console.log(`✅ Prediction received in ${elapsed}s`);
-    console.log(`\n${'='.repeat(80)}`);
-    console.log(`📊 FULL PREDICTION JSON:`);
-    console.log('='.repeat(80));
-    console.log(JSON.stringify(response.data, null, 2));
-
-    // Summary
-    const pkg = response.data;
-    if (pkg.main_pick) {
-      console.log(`\n${'='.repeat(80)}`);
-      console.log(`🎯 SUMMARY:`);
-      console.log('='.repeat(80));
-      console.log(`Main Pick: ${pkg.main_pick.market} → ${pkg.main_pick.pick}`);
-      console.log(`Confidence: ${pkg.main_pick.confidence}%`);
-      console.log(`Odds: ${pkg.main_pick.odds}`);
-      console.log(`Bet Grade: ${pkg.main_pick.bet_grade}`);
-      console.log(`Edge: ${pkg.main_pick.edge || 'N/A'}`);
-
-      if (pkg.value_pick) {
-        console.log(`\nValue Pick: ${pkg.value_pick.market} → ${pkg.value_pick.pick}`);
-        console.log(`Confidence: ${pkg.value_pick.confidence}%`);
-        console.log(`Odds: ${pkg.value_pick.odds}`);
-      }
-
-      if (pkg.bet_advice) {
-        console.log(`\n💡 Bet Advice:`);
-        console.log(`   Playable: ${pkg.bet_advice.playable}`);
-        console.log(`   Stake: ${pkg.bet_advice.suggested_stake_units} units`);
-        console.log(`   Reason: ${pkg.bet_advice.reason}`);
-      }
-
-      if (pkg.score_prediction) {
-        console.log(`\n⚽ Score Prediction:`);
-        console.log(`   FT: ${pkg.score_prediction.ft}`);
-        console.log(`   HT: ${pkg.score_prediction.ht}`);
-        console.log(`   xG: ${pkg.score_prediction.xg_home} - ${pkg.score_prediction.xg_away}`);
-      }
-
-      if (pkg.risk) {
-        console.log(`\n⚠️ Risk Level: ${pkg.risk.level} (${pkg.risk.score})`);
-        if (pkg.risk.warnings?.length > 0) {
-          console.log(`   Warnings: ${pkg.risk.warnings.join(', ')}`);
-        }
-      }
-
-      if (pkg.ai_commentary) {
-        console.log(`\n💬 AI Commentary:`);
-        console.log(`   ${pkg.ai_commentary}`);
-      }
-    }
-
-    return response.data;
-
-  } catch (error: any) {
-    console.error(`❌ Error for match ${matchId}:`);
-    if (error.response) {
-      console.error(`   Status: ${error.response.status}`);
-      console.error(`   Data: ${JSON.stringify(error.response.data, null, 2)}`);
-    } else {
-      console.error(`   Message: ${error.message}`);
-    }
-    return null;
-  }
-}
-
-async function main() {
-  console.log('🚀 VQWEN v3 Prediction Engine - Batch Analysis');
-  console.log(`📡 AI Engine: ${process.env.AI_ENGINE_URL || 'http://localhost:8007'}`);
-  console.log(`🎯 Matches: ${matchIds.length}`);
-
-  const results: { matchId: string; success: boolean }[] = [];
-
-  for (const matchId of matchIds) {
-    const result = await getPrediction(matchId);
-    if (result) {
-      results.push({ matchId, success: true });
-    } else {
-      results.push({ matchId, success: false });
-    }
-
-    // Small delay between requests
-    await new Promise(resolve => setTimeout(resolve, 1000));
-  }
-
-  console.log(`\n${'='.repeat(80)}`);
-  console.log(`📊 BATCH SUMMARY:`);
-  console.log('='.repeat(80));
-  results.forEach((r, i) => {
-    console.log(`${r.success ? '✅' : '❌'} ${i + 1}. ${r.matchId}`);
-  });
-  console.log(`\nTotal: ${results.filter(r => r.success).length}/${results.length} successful`);
-
-  await prisma.$disconnect();
-}
-
-main().catch(console.error);

scripts/warm-image-cache.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+# Warms the R2 image bucket by requesting every known team / competition /
+# country image once through the image-proxy Worker, which mirrors each one
+# into R2 (see workers/image-proxy).
+#
+# Run on the production server (needs docker access to iddaai-postgres):
+#   ./warm-image-cache.sh https://files.example.com
+set -euo pipefail
+
+BASE_URL="${1:?Usage: $0 <image-base-url>}"
+BASE_URL="${BASE_URL%/}"
+PSQL=(docker exec iddaai-postgres psql -U iddaai_user -d iddaai_db -At -c)
+
+{
+  "${PSQL[@]}" "SELECT 'teams/' || id FROM teams"
+  "${PSQL[@]}" "SELECT 'competitions/' || id FROM leagues"
+  "${PSQL[@]}" "SELECT 'areas/' || id FROM countries"
+} | xargs -P 8 -I{} curl -sS -o /dev/null -w "%{http_code}\n" "$BASE_URL/{}" \
+  | sort | uniq -c \
+  | awk '{printf "HTTP %s: %s istek\n", $2, $1}'

src/app.module.ts

@@ -46,6 +46,7 @@ import { SocialPosterModule } from "./modules/social-poster/social-poster.module
 // Sports Domain Modules
 import { MatchesModule } from "./modules/matches/matches.module";
 import { PredictionsModule } from "./modules/predictions/predictions.module";
+import { ValueBoardModule } from "./modules/value-board/value-board.module";
 import { LeaguesModule } from "./modules/leagues/leagues.module";
 import { AnalysisModule } from "./modules/analysis/analysis.module";
 import { CouponsModule } from "./modules/coupons/coupons.module";
@@ -200,6 +201,7 @@ const historicalFeederMode = process.env.FEEDER_MODE === "historical";
     // Sports Domain Modules
     MatchesModule,
     PredictionsModule,
+    ValueBoardModule,
     LeaguesModule,
     AnalysisModule,
     CouponsModule,

src/common/utils/image-url.util.ts

@@ -0,0 +1,35 @@
+/**
+ * Central builder for team / competition / country image URLs.
+ *
+ * Images are served from a Cloudflare R2 bucket fronted by the
+ * `workers/image-proxy` Worker, which lazily mirrors each image from the
+ * upstream provider (file.mackolikfeeds.com) into R2 on first request.
+ *
+ * Set IMAGE_BASE_URL (no trailing slash, e.g. https://files.example.com)
+ * to serve from the Worker. When unset, URLs point directly at the
+ * upstream provider so behaviour is unchanged until the bucket is live.
+ */
+const DEFAULT_IMAGE_BASE_URL = "https://file.mackolikfeeds.com";
+
+function imageBaseUrl(): string {
+  return (
+    process.env.IMAGE_BASE_URL?.replace(/\/+$/, "") || DEFAULT_IMAGE_BASE_URL
+  );
+}
+
+export function teamLogoUrl(teamId?: string | null): string | undefined {
+  if (!teamId) return undefined;
+  return `${imageBaseUrl()}/teams/${teamId}`;
+}
+
+export function competitionLogoUrl(
+  competitionId?: string | null,
+): string | undefined {
+  if (!competitionId) return undefined;
+  return `${imageBaseUrl()}/competitions/${competitionId}`;
+}
+
+export function countryFlagUrl(countryId?: string | null): string | undefined {
+  if (!countryId) return undefined;
+  return `${imageBaseUrl()}/areas/${countryId}`;
+}

src/common/utils/image.util.ts

@@ -1,59 +0,0 @@
-import { existsSync, createWriteStream, mkdirSync } from "fs";
-import { dirname } from "path";
-import axios from "axios";
-import { Logger } from "@nestjs/common";
-
-export class ImageUtils {
-  private static readonly logger = new Logger("ImageUtils");
-
-  /**
-   * Downloads an image from a URL and saves it to a local path.
-   * Skips download if file already exists.
-   */
-  static async downloadImage(url: string, localPath: string): Promise<boolean> {
-    try {
-      // Check if file exists
-      if (existsSync(localPath)) {
-        return true;
-      }
-
-      // Ensure directory exists
-      const dir = dirname(localPath);
-      if (!existsSync(dir)) {
-        mkdirSync(dir, { recursive: true });
-      }
-
-      // Download
-      const response = await axios({
-        url,
-        method: "GET",
-        responseType: "stream",
-        timeout: 5000,
-        validateStatus: (status) => status === 200, // Only save if 200 OK
-      });
-
-      const writer = createWriteStream(localPath);
-
-      response.data.pipe(writer);
-
-      return new Promise((resolve, reject) => {
-        writer.on("finish", () => resolve(true));
-        writer.on("error", (err) => {
-          this.logger.warn(
-            `Failed to write image to ${localPath}: ${err.message}`,
-          );
-          reject(new Error(`Failed to write image to ${localPath}`));
-        });
-      });
-    } catch (error: any) {
-      // Log warning but don't break the application
-      // 404s are common for missing logos
-      if (error.response?.status !== 404) {
-        this.logger.warn(
-          `Failed to download image from ${url}: ${error.message}`,
-        );
-      }
-      return false;
-    }
-  }
-}

src/modules/admin/admin.controller.ts

@@ -394,4 +394,202 @@ export class AdminController {
       predictions: totalPredictions,
     });
   }
+
+  // ================== Model Performance (Forward-Test) ==================
+
+  @Get("model-performance")
+  @ApiOperation({
+    summary:
+      "Per-market calibration (model% vs actual%), ROI and decision rationale " +
+      "from settled prediction_runs. Powers the admin Model Performance page.",
+  })
+  @SwaggerResponse({ status: 200, schema: { type: "object" } })
+  async getModelPerformance(
+    @Query("days") daysRaw?: string,
+  ): Promise<ApiResponse<ModelPerformanceResult>> {
+    const days = Math.min(Math.max(Number(daysRaw) || 90, 1), 1000);
+    const sinceMs = Date.now() - days * 24 * 60 * 60 * 1000;
+
+    // Pull settled rows in window. markets_settled holds one entry per market.
+    const rows = await this.prisma.$queryRawUnsafe<
+      Array<{ payload_summary: unknown; generated_at: Date }>
+    >(
+      `
+        SELECT pr.payload_summary, pr.generated_at
+        FROM prediction_runs pr
+        WHERE pr.eventual_outcome IS NOT NULL
+          AND pr.generated_at >= $1
+          AND pr.payload_summary -> 'settlement' -> 'markets_settled' IS NOT NULL
+        ORDER BY pr.generated_at DESC
+        LIMIT 50000
+      `,
+      new Date(sinceMs),
+    );
+
+    // ── Aggregate per market ──────────────────────────────────────────
+    type Acc = {
+      market: string;
+      n: number;
+      wins: number;
+      sumShown: number; // Σ shown_confidence (0–100)
+      shownCount: number;
+      // 10-bin reliability for ECE
+      bins: Array<{ sumP: number; sumY: number; n: number }>;
+      // betting (only playable BET rows with odds)
+      betN: number;
+      betWins: number;
+      betProfit: number;
+      betStake: number;
+      // rationale tally
+      actions: Record<string, number>;
+      tiers: Record<string, number>;
+    };
+    const acc = new Map<string, Acc>();
+    const ensure = (mk: string): Acc => {
+      let a = acc.get(mk);
+      if (!a) {
+        a = {
+          market: mk,
+          n: 0,
+          wins: 0,
+          sumShown: 0,
+          shownCount: 0,
+          bins: Array.from({ length: 10 }, () => ({ sumP: 0, sumY: 0, n: 0 })),
+          betN: 0,
+          betWins: 0,
+          betProfit: 0,
+          betStake: 0,
+          actions: {},
+          tiers: {},
+        };
+        acc.set(mk, a);
+      }
+      return a;
+    };
+
+    let totalSettledMarkets = 0;
+    for (const row of rows) {
+      const summary =
+        row.payload_summary && typeof row.payload_summary === "object"
+          ? (row.payload_summary as Record<string, unknown>)
+          : {};
+      const settlement =
+        summary.settlement && typeof summary.settlement === "object"
+          ? (summary.settlement as Record<string, unknown>)
+          : {};
+      const markets = Array.isArray(settlement.markets_settled)
+        ? (settlement.markets_settled as Array<Record<string, unknown>>)
+        : [];
+
+      for (const m of markets) {
+        const market = typeof m.market === "string" ? m.market : "";
+        if (!market) continue;
+        const won = m.won === true;
+        const shown =
+          m.shown_confidence != null ? Number(m.shown_confidence) : null;
+        const a = ensure(market);
+        a.n += 1;
+        if (won) a.wins += 1;
+        totalSettledMarkets += 1;
+
+        if (shown != null && Number.isFinite(shown)) {
+          a.sumShown += shown;
+          a.shownCount += 1;
+          const p = Math.min(Math.max(shown / 100, 0), 0.999999);
+          const bi = Math.min(9, Math.floor(p * 10));
+          a.bins[bi].sumP += p;
+          a.bins[bi].sumY += won ? 1 : 0;
+          a.bins[bi].n += 1;
+        }
+
+        const odds = m.odds != null ? Number(m.odds) : null;
+        const isBet = m.playable === true && m.action === "BET";
+        if (isBet && odds != null && Number.isFinite(odds) && odds > 1.01) {
+          a.betN += 1;
+          if (won) {
+            a.betWins += 1;
+            a.betProfit += odds - 1;
+          } else {
+            a.betProfit -= 1;
+          }
+          a.betStake += 1;
+        }
+
+        const action = typeof m.action === "string" ? m.action : "—";
+        a.actions[action] = (a.actions[action] ?? 0) + 1;
+        const tier = typeof m.value_tier === "string" ? m.value_tier : "—";
+        a.tiers[tier] = (a.tiers[tier] ?? 0) + 1;
+      }
+    }
+
+    const markets = Array.from(acc.values())
+      .map((a) => {
+        const actualPct = a.n > 0 ? (a.wins / a.n) * 100 : 0;
+        const shownPct = a.shownCount > 0 ? a.sumShown / a.shownCount : 0;
+        // ECE: Σ |acc_bin - conf_bin| * (n_bin / N)
+        let ece = 0;
+        for (const b of a.bins) {
+          if (b.n === 0) continue;
+          const conf = b.sumP / b.n;
+          const acc2 = b.sumY / b.n;
+          ece += Math.abs(acc2 - conf) * (b.n / a.shownCount || 0);
+        }
+        return {
+          market: a.market,
+          samples: a.n,
+          shown_pct: Number(shownPct.toFixed(1)),
+          actual_pct: Number(actualPct.toFixed(1)),
+          gap: Number((shownPct - actualPct).toFixed(1)),
+          ece: Number((ece * 100).toFixed(1)),
+          calibration: (Math.abs(shownPct - actualPct) <= 4
+            ? "good"
+            : shownPct > actualPct
+              ? "overconfident"
+              : "underconfident") as
+            | "good"
+            | "overconfident"
+            | "underconfident",
+          bet_count: a.betN,
+          bet_hit_pct:
+            a.betN > 0 ? Number(((a.betWins / a.betN) * 100).toFixed(1)) : 0,
+          bet_roi_pct:
+            a.betStake > 0
+              ? Number(((a.betProfit / a.betStake) * 100).toFixed(1))
+              : 0,
+          actions: a.actions,
+          tiers: a.tiers,
+        };
+      })
+      .sort((x, y) => y.samples - x.samples);
+
+    const result: ModelPerformanceResult = {
+      window_days: days,
+      settled_runs: Number(rows.length),
+      settled_markets: totalSettledMarkets,
+      generated_at: new Date().toISOString(),
+      markets,
+    };
+    return createSuccessResponse(result);
+  }
+}
+
+interface ModelPerformanceResult {
+  window_days: number;
+  settled_runs: number;
+  settled_markets: number;
+  generated_at: string;
+  markets: Array<{
+    market: string;
+    samples: number;
+    shown_pct: number;
+    actual_pct: number;
+    gap: number;
+    ece: number;
+    calibration: "good" | "overconfident" | "underconfident";
+    bet_count: number;
+    bet_hit_pct: number;
+    bet_roi_pct: number;
+    actions: Record<string, number>;
+    tiers: Record<string, number>;
+  }>;
 }

src/modules/coupons/services/smart-coupon.service.ts

@@ -21,6 +21,8 @@ export interface PredictionPickRow {
   odds: number;
   raw_confidence: number;
   calibrated_confidence: number;
+  unified_score: number;
+  unified_score_label: 'very_reliable' | 'reliable' | 'moderate' | 'low';
   min_required_confidence: number;
   edge: number;
   play_score: number;
@@ -35,6 +37,8 @@ export interface PredictionBetSummaryRow {
   pick: string;
   raw_confidence: number;
   calibrated_confidence: number;
+  unified_score: number;
+  unified_score_label: 'very_reliable' | 'reliable' | 'moderate' | 'low';
   bet_grade: BetGrade;
   playable: boolean;
   stake_units: number;

src/modules/feeder/feeder-persistence.service.ts

@@ -22,7 +22,6 @@ import {
   DbMarketPayload,
   BasketballTeamStats,
 } from "./feeder.types";
-import { ImageUtils } from "../../common/utils/image.util";
 import { deriveStoredMatchStatus } from "../../common/utils/match-status.util";
 
 @Injectable()
@@ -164,24 +163,6 @@ export class FeederPersistenceService {
     oddsArray: DbMarketPayload[],
     officialsData: MatchOfficial[],
   ): Promise<boolean> {
-    // START IMAGE DOWNLOADS (NON-BLOCKING)
-    const imageDownloads: Promise<void>[] = [];
-
-    const leagueId = this.safeString(league.id);
-    if (leagueId) {
-      const logoUrl = `https://file.mackolikfeeds.com/competitions/${leagueId}`;
-      const localPath = `public/uploads/competitions/${leagueId}.png`;
-      imageDownloads.push(
-        ImageUtils.downloadImage(logoUrl, localPath)
-          .then(() => void 0)
-          .catch((err) => {
-            this.logger.error(
-              `Failed to download league logo ${leagueId}: ${err}`,
-            );
-          }),
-      );
-    }
-
     const teamsToUpsert = [
       {
         id: homeTeamId,
@@ -197,20 +178,6 @@ export class FeederPersistenceService {
       },
     ];
 
-    for (const team of teamsToUpsert) {
-      const teamLogoUrl = `https://file.mackolikfeeds.com/teams/${team.id}`;
-      const teamLocalPath = `public/uploads/teams/${team.id}.png`;
-      imageDownloads.push(
-        ImageUtils.downloadImage(teamLogoUrl, teamLocalPath)
-          .then(() => void 0)
-          .catch((err) => {
-            this.logger.error(
-              `Failed to download team logo ${team.id}: ${err}`,
-            );
-          }),
-      );
-    }
-
     // DATABASE TRANSACTION
     try {
       await this.prisma.$transaction(
@@ -264,7 +231,6 @@ export class FeederPersistenceService {
                   countryId: countryId,
                   sport: sport,
                   competitionSlug: league.competitionSlug,
-                  logoUrl: `/uploads/competitions/${finalLeagueId}.png`,
                 } as any,
               });
               if (league.sortOrder !== undefined) {
@@ -291,10 +257,7 @@ export class FeederPersistenceService {
 
           if (teamsToCreate.length > 0) {
             await tx.team.createMany({
-              data: teamsToCreate.map((t) => ({
-                ...t,
-                logoUrl: `/uploads/teams/${t.id}.png`,
-              })),
+              data: teamsToCreate,
               skipDuplicates: true,
             });
           }
@@ -304,7 +267,6 @@ export class FeederPersistenceService {
               where: { id: team.id },
               data: {
                 name: team.name,
-                logoUrl: `/uploads/teams/${team.id}.png`,
               },
             });
           }
@@ -614,9 +576,6 @@ export class FeederPersistenceService {
         { maxWait: 40000, timeout: 40000 },
       );
 
-      // WAIT FOR IMAGES AFTER TRANSACTION
-      await Promise.allSettled(imageDownloads);
-
       this.logger.log(`✅ SAVED: [${matchId}] ${matchSummary.matchName}`);
       return true;
     } catch (error: any) {

src/modules/leagues/leagues.service.ts

@@ -1,6 +1,10 @@
 import { Injectable, Logger } from "@nestjs/common";
 import { PrismaService } from "../../database/prisma.service";
 import { Sport } from "@prisma/client";
+import {
+  countryFlagUrl,
+  teamLogoUrl,
+} from "../../common/utils/image-url.util";
 
 @Injectable()
 export class LeaguesService {
@@ -12,19 +16,21 @@ export class LeaguesService {
    * Get all countries
    */
   async findAllCountries() {
-    return this.prisma.country.findMany({
+    const countries = await this.prisma.country.findMany({
       orderBy: { name: "asc" },
     });
+    return countries.map((c) => ({ ...c, flag: countryFlagUrl(c.id) }));
   }
 
   /**
    * Get country by ID
    */
   async findCountryById(id: string) {
-    return this.prisma.country.findUnique({
+    const country = await this.prisma.country.findUnique({
       where: { id },
       include: { leagues: true },
     });
+    return country ? { ...country, flag: countryFlagUrl(country.id) } : null;
   }
 
   /**
@@ -66,7 +72,7 @@ export class LeaguesService {
    * Get all teams
    */
   async findAllTeams(sport?: Sport, search?: string) {
-    return this.prisma.team.findMany({
+    const teams = await this.prisma.team.findMany({
       where: {
         ...(sport ? { sport } : {}),
         ...(search ? { name: { contains: search, mode: "insensitive" } } : {}),
@@ -74,28 +80,31 @@ export class LeaguesService {
       orderBy: { name: "asc" },
       take: 100,
     });
+    return teams.map((t) => ({ ...t, logo: teamLogoUrl(t.id) }));
   }
 
   /**
    * Get team by ID
    */
   async findTeamById(id: string) {
-    return this.prisma.team.findUnique({
+    const team = await this.prisma.team.findUnique({
       where: { id },
     });
+    return team ? { ...team, logo: teamLogoUrl(team.id) } : null;
   }
 
   /**
    * Search teams by name
    */
   async searchTeams(name: string, sport?: Sport) {
-    return this.prisma.team.findMany({
+    const teams = await this.prisma.team.findMany({
       where: {
         name: { contains: name, mode: "insensitive" },
         ...(sport ? { sport } : {}),
       },
       take: 20,
     });
+    return teams.map((t) => ({ ...t, logo: teamLogoUrl(t.id) }));
   }
 
   /**
@@ -161,13 +170,9 @@ export class LeaguesService {
         status: m.status,
         state: m.state,
         homeTeamName: m.homeTeam?.name,
-        homeTeamLogo: m.homeTeamId
-          ? `https://file.mackolikfeeds.com/teams/${m.homeTeamId}`
-          : null,
+        homeTeamLogo: teamLogoUrl(m.homeTeamId) ?? null,
         awayTeamName: m.awayTeam?.name,
-        awayTeamLogo: m.awayTeamId
-          ? `https://file.mackolikfeeds.com/teams/${m.awayTeamId}`
-          : null,
+        awayTeamLogo: teamLogoUrl(m.awayTeamId) ?? null,
         leagueName: m.league?.name,
         countryName: m.league?.country?.name,
       })),

src/modules/matches/matches.controller.ts

@@ -111,6 +111,22 @@ export class MatchesController {
     return this.matchesService.getActiveLeagues(sport || Sport.FOOTBALL);
   }
 
+  /**
+   * GET /matches/:id/odds-movement
+   * Opening→closing odds movement per market/selection (from live_odds_history)
+   */
+  @Public()
+  @Get(":id/odds-movement")
+  @ApiOperation({ summary: "Opening→closing odds movement for a match" })
+  @ApiParam({ name: "id", description: "Match ID" })
+  @ApiResponse({ status: 200, description: "{ market: { selection: { open, close } } }" })
+  async getOddsMovement(@Param("id") id: string) {
+    if (!id) {
+      throw new BadRequestException("Match ID is required");
+    }
+    return this.matchesService.getOddsMovement(id);
+  }
+
   /**
    * GET /matches/:id
    * Get full match details

src/modules/matches/matches.service.ts

@@ -16,6 +16,10 @@ import {
   LIVE_STATUS_VALUES_FOR_DB,
   getDisplayMatchStatus,
 } from "../../common/utils/match-status.util";
+import {
+  countryFlagUrl,
+  teamLogoUrl,
+} from "../../common/utils/image-url.util";
 
 @Injectable()
 export class MatchesService {
@@ -28,6 +32,51 @@ export class MatchesService {
     this.loadTopLeagues();
   }
 
+  /**
+   * Per-match odds movement (opening→closing) from live_odds_history.
+   * Returns { [market]: { [selection]: { open, close } } } with the same
+   * Turkish market/selection labels used in match.odds, so the UI can line
+   * them up directly. Returns {} if there is no data or the table is absent.
+   */
+  async getOddsMovement(
+    matchId: string,
+  ): Promise<Record<string, Record<string, { open: number; close: number }>>> {
+    try {
+      const rows = await this.prisma.$queryRawUnsafe<
+        Array<{
+          market: string;
+          selection: string;
+          open: number | null;
+          close: number | null;
+        }>
+      >(
+        `SELECT market, selection,
+                (array_agg(new_value ORDER BY change_time ASC))[1]  AS open,
+                (array_agg(new_value ORDER BY change_time DESC))[1] AS close
+         FROM live_odds_history
+         WHERE match_id = $1
+         GROUP BY market, selection`,
+        matchId,
+      );
+      const out: Record<
+        string,
+        Record<string, { open: number; close: number }>
+      > = {};
+      for (const r of rows) {
+        if (r.open == null || r.close == null) continue;
+        (out[r.market] ??= {})[r.selection] = {
+          open: Number(r.open),
+          close: Number(r.close),
+        };
+      }
+      return out;
+    } catch (err) {
+      const msg = err instanceof Error ? err.message : String(err);
+      this.logger.warn(`getOddsMovement failed for ${matchId}: ${msg}`);
+      return {};
+    }
+  }
+
   private loadTopLeagues() {
     try {
       const filePath = path.join(process.cwd(), "top_leagues.json");
@@ -53,20 +102,12 @@ export class MatchesService {
     }
   }
 
-  /**
-   * Generate URL for the country flag served from Mackolik
-   */
   private getCountryFlagUrl(countryId?: string | null): string | undefined {
-    if (!countryId) return undefined;
-    return `https://file.mackolikfeeds.com/areas/${countryId}`;
+    return countryFlagUrl(countryId);
   }
 
-  /**
-   * Generate URL for the team logo served from local uploads
-   */
   private getTeamLogoUrl(teamId?: string | null): string | undefined {
-    if (!teamId) return undefined;
-    return `https://file.mackolikfeeds.com/teams/${teamId}`;
+    return teamLogoUrl(teamId);
   }
 
   private getLiveFilter(): Prisma.LiveMatchWhereInput {

src/modules/predictions/dto/index.ts

@@ -27,6 +27,20 @@ export class MatchInfoDto {
   @ApiProperty({ required: false, default: false })
   is_top_league?: boolean;
 
+  @ApiProperty({
+    required: false,
+    nullable: true,
+    description:
+      "Backtest-derived per-league confidence (ROI + sample size). " +
+      "null when the league has too little data to judge.",
+  })
+  league_confidence?: {
+    label: "high" | "medium" | "low";
+    bet_roi: number;
+    bet_n: number;
+    hit: number;
+  } | null;
+
   @ApiProperty({
     required: false,
     enum: ["football", "basketball"],

src/modules/predictions/predictions.service.ts

@@ -1525,6 +1525,37 @@ export class PredictionsService implements OnModuleInit, OnModuleDestroy {
     payload: MatchPredictionDto,
   ): Promise<void> {
     try {
+      // Finished-match re-analyses (manual validation runs) must not pollute
+      // the forward track record: they would bias settlement ROI, the
+      // per-league karne and engine-version comparisons. Tag them into their
+      // own engine_version bucket so every GROUP BY engine_version isolates
+      // them automatically — the data is kept, the live karne stays clean.
+      const auditMatch = await this.prisma.match.findUnique({
+        where: { id: matchId },
+        select: {
+          state: true,
+          status: true,
+          scoreHome: true,
+          scoreAway: true,
+          mstUtc: true,
+        },
+      });
+      const kickoffMs =
+        auditMatch?.mstUtc != null ? Number(auditMatch.mstUtc) : null;
+      const kickoffLongPast =
+        kickoffMs !== null && Date.now() - kickoffMs > 3 * 60 * 60 * 1000;
+      const isCompletedRun =
+        isMatchCompleted({
+          state: auditMatch?.state ?? null,
+          status: auditMatch?.status ?? null,
+          scoreHome: auditMatch?.scoreHome,
+          scoreAway: auditMatch?.scoreAway,
+        }) || kickoffLongPast;
+      const baseVersion = String(payload.model_version || "unknown");
+      const engineVersion = isCompletedRun
+        ? `${baseVersion}.sim-finished`
+        : baseVersion;
+
       const oddsSnapshot = await this.getPredictionOddsSnapshot(matchId);
       const payloadSummary = this.buildPredictionPayloadSummary(payload);
       await this.prisma.$executeRawUnsafe(
@@ -1539,7 +1570,7 @@ export class PredictionsService implements OnModuleInit, OnModuleDestroy {
           VALUES ($1, $2, $3, $4::jsonb, $5::jsonb)
         `,
         matchId,
-        String(payload.model_version || "unknown"),
+        engineVersion,
         typeof payload.decision_trace_id === "string"
           ? payload.decision_trace_id
           : null,
@@ -1611,7 +1642,69 @@ export class PredictionsService implements OnModuleInit, OnModuleDestroy {
         }))
       : [];
 
+    // ── Forward-test capture (V31e) ────────────────────────────────────
+    // Persist EVERY market's probability + the rationale behind each pick so
+    // the settlement job can later score each market against reality and the
+    // admin "Model Performance" page can show per-market calibration
+    // (model% → actual%) and decision reasons. Compact projection only.
+    // Some runtime fields (betting_brain, is_underdog_reference) are present
+    // in the AI payload but not declared on the DTO — read them via a cast.
+    const marketsFull = Array.isArray(payload.bet_summary)
+      ? payload.bet_summary.map((item) => {
+          const loose = item as unknown as Record<string, unknown>;
+          const bb = (loose.betting_brain ?? {}) as Record<string, unknown>;
+          return {
+            market: item.market,
+            pick: item.pick,
+            odds: item.odds ?? null,
+            model_probability: item.model_probability ?? null,
+            calibrated_confidence: item.calibrated_confidence ?? null,
+            raw_confidence: item.raw_confidence ?? null,
+            calibrated_probability: item.calibrated_probability ?? null,
+            implied_prob: item.implied_prob ?? null,
+            ev_edge: item.ev_edge ?? 0,
+            playable: item.playable ?? false,
+            bet_grade: item.bet_grade ?? "PASS",
+            signal_tier: item.signal_tier ?? null,
+            stake_units: item.stake_units ?? 0,
+            is_underdog_reference: Boolean(loose.is_underdog_reference),
+            action: (bb.action as string | undefined) ?? null,
+            value_tier: (bb.value_tier as string | undefined) ?? null,
+            model_market_gap: (bb.model_market_gap as number | undefined) ?? null,
+            trap_market_flag: Boolean(bb.trap_market_flag),
+            vetoes: Array.isArray(bb.vetoes)
+              ? (bb.vetoes as unknown[]).slice(0, 6)
+              : [],
+            positives: Array.isArray(bb.positives)
+              ? (bb.positives as unknown[]).slice(0, 6)
+              : [],
+            reasons: Array.isArray(item.reasons) ? item.reasons.slice(0, 6) : [],
+          };
+        })
+      : [];
+
+    // Per-outcome probability distribution for each market (graph bars).
+    const marketBoardProbs =
+      payload.market_board && typeof payload.market_board === "object"
+        ? Object.fromEntries(
+            Object.entries(
+              payload.market_board as Record<string, { probs?: unknown }>,
+            ).map(([mkt, entry]) => [
+              mkt,
+              entry && typeof entry === "object" ? (entry.probs ?? null) : null,
+            ]),
+          )
+        : {};
+
     return {
+      markets_full: marketsFull,
+      market_board_probs: marketBoardProbs,
+      betting_brain_version:
+        (
+          (payload as unknown as Record<string, unknown>).betting_brain as
+            | { version?: string }
+            | undefined
+        )?.version ?? null,
       model_version: payload.model_version,
       calibration_version: payload.calibration_version ?? null,
       shadow_engine_version: payload.shadow_engine_version ?? null,

src/modules/social-poster/assets/basketball.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"><circle fill="#F4900C" cx="18" cy="18" r="18"/><path fill="#231F20" d="M36 17h-8.981c.188-5.506 1.943-9.295 4.784-10.546-.445-.531-.926-1.027-1.428-1.504-2.83 1.578-5.145 5.273-5.354 12.049H19V0h-2v17h-6.021c-.208-6.776-2.523-10.471-5.353-12.049-.502.476-.984.972-1.428 1.503C7.039 7.705 8.793 11.494 8.981 17H0v2h8.981c-.188 5.506-1.942 9.295-4.783 10.546.445.531.926 1.027 1.428 1.504 2.831-1.578 5.145-5.273 5.353-12.05H17v17h2V19h6.021c.209 6.776 2.523 10.471 5.354 12.05.502-.476.984-.973 1.428-1.504-2.841-1.251-4.595-5.04-4.784-10.546H36v-2z"/></svg>

src/modules/social-poster/assets/flag-de.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"><path fill="#FFCD05" d="M0 27c0 2.209 1.791 4 4 4h28c2.209 0 4-1.791 4-4v-4H0v4z"/><path fill="#ED1F24" d="M0 14h36v9H0z"/><path fill="#141414" d="M32 5H4C1.791 5 0 6.791 0 9v5h36V9c0-2.209-1.791-4-4-4z"/></svg>

src/modules/social-poster/assets/flag-es.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"><path fill="#C60A1D" d="M36 27c0 2.209-1.791 4-4 4H4c-2.209 0-4-1.791-4-4V9c0-2.209 1.791-4 4-4h28c2.209 0 4 1.791 4 4v18z"/><path fill="#FFC400" d="M0 12h36v12H0z"/><path fill="#EA596E" d="M9 17v3c0 1.657 1.343 3 3 3s3-1.343 3-3v-3H9z"/><path fill="#F4A2B2" d="M12 16h3v3h-3z"/><path fill="#DD2E44" d="M9 16h3v3H9z"/><ellipse fill="#EA596E" cx="12" cy="14.5" rx="3" ry="1.5"/><ellipse fill="#FFAC33" cx="12" cy="13.75" rx="3" ry=".75"/><path fill="#99AAB5" d="M7 16h1v7H7zm9 0h1v7h-1z"/><path fill="#66757F" d="M6 22h3v1H6zm9 0h3v1h-3zm-8-7h1v1H7zm9 0h1v1h-1z"/></svg>

src/modules/social-poster/assets/flag-eu.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"><path fill="#039" d="M32 5H4C1.791 5 0 6.791 0 9v18c0 2.209 1.791 4 4 4h28c2.209 0 4-1.791 4-4V9c0-2.209-1.791-4-4-4z"/><path d="M18.539 9.705l.849-.617h-1.049l-.325-.998-.324.998h-1.049l.849.617-.325.998.849-.617.849.617zm0 17.333l.849-.617h-1.049l-.325-.998-.324.998h-1.049l.849.617-.325.998.849-.617.849.617zm-8.666-8.667l.849-.617h-1.05l-.324-.998-.325.998H7.974l.849.617-.324.998.849-.617.849.617zm1.107-4.285l.849-.617h-1.05l-.324-.998-.324.998h-1.05l.849.617-.324.998.849-.617.849.617zm0 8.619l.849-.617h-1.05l-.324-.998-.324.998h-1.05l.849.617-.324.998.849-.617.849.617zm3.226-11.839l.849-.617h-1.05l-.324-.998-.324.998h-1.05l.849.617-.324.998.849-.617.849.617zm0 15.067l.849-.617h-1.05l-.324-.998-.324.998h-1.05l.849.617-.324.998.849-.616.849.616zm11.921-7.562l-.849-.617h1.05l.324-.998.325.998h1.049l-.849.617.324.998-.849-.617-.849.617zm-1.107-4.285l-.849-.617h1.05l.324-.998.324.998h1.05l-.849.617.324.998-.849-.617-.849.617zm0 8.619l-.849-.617h1.05l.324-.998.324.998h1.05l-.849.617.324.998-.849-.617-.849.617zm-3.226-11.839l-.849-.617h1.05l.324-.998.324.998h1.05l-.849.617.324.998-.849-.617-.849.617zm0 15.067l-.849-.617h1.05l.324-.998.324.998h1.05l-.849.617.324.998-.849-.616-.849.616z" fill="#FC0"/></svg>

src/modules/social-poster/assets/flag-fr.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"><path fill="#ED2939" d="M36 27c0 2.209-1.791 4-4 4h-8V5h8c2.209 0 4 1.791 4 4v18z"/><path fill="#002495" d="M4 5C1.791 5 0 6.791 0 9v18c0 2.209 1.791 4 4 4h8V5H4z"/><path fill="#EEE" d="M12 5h12v26H12z"/></svg>

src/modules/social-poster/assets/flag-gb.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"><path fill="#00247D" d="M0 9.059V13h5.628zM4.664 31H13v-5.837zM23 25.164V31h8.335zM0 23v3.941L5.63 23zM31.337 5H23v5.837zM36 26.942V23h-5.631zM36 13V9.059L30.371 13zM13 5H4.664L13 10.837z"/><path fill="#CF1B2B" d="M25.14 23l9.712 6.801c.471-.479.808-1.082.99-1.749L28.627 23H25.14zM13 23h-2.141l-9.711 6.8c.521.53 1.189.909 1.938 1.085L13 23.943V23zm10-10h2.141l9.711-6.8c-.521-.53-1.188-.909-1.937-1.085L23 12.057V13zm-12.141 0L1.148 6.2C.677 6.68.34 7.282.157 7.949L7.372 13h3.487z"/><path fill="#EEE" d="M36 21H21v10h2v-5.836L31.335 31H32c1.117 0 2.126-.461 2.852-1.199L25.14 23h3.487l7.215 5.052c.093-.337.158-.686.158-1.052v-.058L30.369 23H36v-2zM0 21v2h5.63L0 26.941V27c0 1.091.439 2.078 1.148 2.8l9.711-6.8H13v.943l-9.914 6.941c.294.07.598.116.914.116h.664L13 25.163V31h2V21H0zM36 9c0-1.091-.439-2.078-1.148-2.8L25.141 13H23v-.943l9.915-6.942C32.62 5.046 32.316 5 32 5h-.663L23 10.837V5h-2v10h15v-2h-5.629L36 9.059V9zM13 5v5.837L4.664 5H4c-1.118 0-2.126.461-2.852 1.2l9.711 6.8H7.372L.157 7.949C.065 8.286 0 8.634 0 9v.059L5.628 13H0v2h15V5h-2z"/><path fill="#CF1B2B" d="M21 15V5h-6v10H0v6h15v10h6V21h15v-6z"/></svg>

src/modules/social-poster/assets/flag-it.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"><path fill="#CE2B37" d="M36 27c0 2.209-1.791 4-4 4h-8V5h8c2.209 0 4 1.791 4 4v18z"/><path fill="#009246" d="M4 5C1.791 5 0 6.791 0 9v18c0 2.209 1.791 4 4 4h8V5H4z"/><path fill="#EEE" d="M12 5h12v26H12z"/></svg>

src/modules/social-poster/assets/flag-tr.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"><path fill="#E30917" d="M36 27c0 2.209-1.791 4-4 4H4c-2.209 0-4-1.791-4-4V9c0-2.209 1.791-4 4-4h28c2.209 0 4 1.791 4 4v18z"/><path fill="#EEE" d="M16 24c-3.314 0-6-2.685-6-6 0-3.314 2.686-6 6-6 1.31 0 2.52.425 3.507 1.138-1.348-1.524-3.312-2.491-5.507-2.491-4.061 0-7.353 3.292-7.353 7.353 0 4.062 3.292 7.354 7.353 7.354 2.195 0 4.16-.967 5.507-2.492C18.521 23.575 17.312 24 16 24zm3.913-5.77l2.44.562.22 2.493 1.288-2.146 2.44.561-1.644-1.888 1.287-2.147-2.303.98-1.644-1.889.22 2.494z"/></svg>