v28

ai-engine/services/single_match_orchestrator.py

@@ -30,12 +30,18 @@ from models.v20_ensemble import FullMatchPrediction
 from models.v25_ensemble import V25Predictor, get_v25_predictor
 from models.v27_predictor import V27Predictor, compute_divergence, compute_value_edge
 from features.odds_band_analyzer import OddsBandAnalyzer
-from models.basketball_v25 import (
-    BasketballMatchPrediction,
-    get_basketball_v25_predictor,
-)
+try:
+    from models.basketball_v25 import (
+        BasketballMatchPrediction,
+        get_basketball_v25_predictor,
+    )
+except ImportError:
+    BasketballMatchPrediction = Any
+    def get_basketball_v25_predictor():
+        raise ImportError("Basketball predictor is not available")
 from core.engines.player_predictor import PlayerPrediction, get_player_predictor
 from services.feature_enrichment import FeatureEnrichmentService
+from services.betting_brain import BettingBrain
 from services.v26_shadow_engine import V26ShadowEngine, get_v26_shadow_engine
 from utils.top_leagues import load_top_league_ids
 from utils.league_reliability import load_league_reliability
@@ -69,6 +75,7 @@ class MatchData:
     substate: Optional[str] = None
     current_score_home: Optional[int] = None
     current_score_away: Optional[int] = None
+    lineup_confidence: float = 0.0
 
 
 class SingleMatchOrchestrator:
@@ -144,7 +151,7 @@ class SingleMatchOrchestrator:
         self.v26_shadow_engine: Optional[V26ShadowEngine] = None
         self.basketball_predictor: Optional[Any] = None
         self.dsn = get_clean_dsn()
-        self.engine_mode = str(os.getenv("AI_ENGINE_MODE", "v25")).strip().lower()
+        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.enrichment = FeatureEnrichmentService()
@@ -527,12 +534,18 @@ class SingleMatchOrchestrator:
         }
 
     def _get_squad_features(self, data: MatchData) -> Dict[str, float]:
-        """Non-fatal squad analysis. Returns zero-defaults on failure."""
+        """Non-fatal squad analysis. Returns neutral-average defaults on failure.
+
+        Design note (V32-fix): Previous 0.0 defaults caused the model to treat
+        missing lineups as 'both teams have zero quality', producing overly
+        conservative predictions (e.g. static 1.5 Under).  Neutral averages let
+        the model fall back on stronger signals (odds, ELO, form, H2H).
+        """
         defaults = {
-            'home_squad_quality': 0.0, 'away_squad_quality': 0.0, 'squad_diff': 0.0,
-            'home_key_players': 0.0, 'away_key_players': 0.0,
+            'home_squad_quality': 0.50, 'away_squad_quality': 0.50, 'squad_diff': 0.0,
+            'home_key_players': 3.0, 'away_key_players': 3.0,
             'home_missing_impact': 0.0, 'away_missing_impact': 0.0,
-            'home_goals_form': 0.0, 'away_goals_form': 0.0,
+            'home_goals_form': 1.3, 'away_goals_form': 1.3,
         }
         try:
             engine = get_player_predictor()
@@ -559,27 +572,186 @@ class SingleMatchOrchestrator:
             print(f"⚠️ Squad features failed: {e}")
             return defaults
 
+    # ── V25 internal key → _build_v25_prediction key mapping ──
+    _V25_KEY_MAP = {
+        "ms": "MS",
+        "ou15": "OU15",
+        "ou25": "OU25",
+        "ou35": "OU35",
+        "btts": "BTTS",
+        "ht_result": "HT",
+        "ht_ou05": "HT_OU05",
+        "ht_ou15": "HT_OU15",
+        "htft": "HTFT",
+        "cards_ou45": "CARDS",
+        "handicap_ms": "HCAP",
+        "odd_even": "OE",
+    }
+
     def _get_v25_signal(
         self,
         data: MatchData,
         features: Optional[Dict[str, float]] = None,
     ) -> Dict[str, Any]:
+        """
+        Get V25 ensemble predictions for all available markets.
+        Returns a dict keyed by UPPERCASE market name (MS, OU25, BTTS, etc.)
+        each with a 'probs' sub-dict that _prob_map can consume.
+
+        CRITICAL: Keys MUST be uppercase to match _build_v25_prediction lookups.
+        """
         v25 = self._get_v25_predictor()
         feature_row = features or self._build_v25_features(data)
-        return v25.predict_market_bundle(
-            features=feature_row,
-            odds=self._sanitize_v25_odds(data.odds_data or {}),
-        )
+
+        signal: Dict[str, Any] = {}
+
+        def _temperature_scale(probs_dict: Dict[str, float], temperature: float = 2.5) -> Dict[str, float]:
+            """
+            Apply temperature scaling to soften overconfident model outputs.
+
+            LightGBM often produces extreme probabilities (e.g., 0.999 / 0.001).
+            Temperature scaling converts to log-odds, divides by T, then re-normalizes.
+            T=1.0 → no change, T>1 → softer probabilities.
+
+            Standard approach for post-hoc model calibration (Guo et al., 2017).
+            """
+            import math
+            eps = 1e-7  # numerical stability
+            n = len(probs_dict)
+
+            # Determine appropriate temperature based on market type
+            # Binary markets (2-class) tend to be more overconfident in LGB
+            if n <= 2:
+                T = max(temperature, 2.0)
+            elif n == 3:
+                T = max(temperature * 0.8, 1.5)  # 3-way slightly less aggressive
+            else:
+                T = max(temperature * 0.6, 1.3)  # 9-way (HTFT) already spread
+
+            # Convert to log-odds and apply temperature
+            labels = list(probs_dict.keys())
+            log_odds = []
+            for label in labels:
+                p = max(eps, min(1.0 - eps, float(probs_dict[label])))
+                log_odds.append(math.log(p) / T)
+
+            # Softmax re-normalization
+            max_lo = max(log_odds)
+            exp_vals = [math.exp(lo - max_lo) for lo in log_odds]
+            total = sum(exp_vals)
+
+            scaled = {}
+            for i, label in enumerate(labels):
+                scaled[label] = exp_vals[i] / total
+
+            return scaled
+
+        def _enrich_signal_entry(probs_dict: Dict[str, float]) -> Dict[str, Any]:
+            """Add pick, probability, confidence to a signal entry from its probs.
+
+            Applies temperature scaling to convert overconfident LightGBM outputs
+            into realistic, calibrated probabilities.
+            """
+            # Apply temperature scaling to soften extreme probabilities
+            scaled_probs = _temperature_scale(probs_dict, temperature=2.5)
+
+            best_label = max(scaled_probs, key=scaled_probs.get)
+            best_prob = float(scaled_probs[best_label])
+            return {
+                "probs": scaled_probs,
+                "raw_probs": probs_dict,  # keep originals for debugging
+                "pick": best_label,
+                "probability": best_prob,
+                "confidence": round(best_prob * 100.0, 1),
+            }
+
+        # Core markets using dedicated methods
+        h, d, a = v25.predict_ms(feature_row)
+        signal["MS"] = _enrich_signal_entry({"1": h, "X": d, "2": a})
+        print(f"  [V25-SIGNAL] MS → H={h:.4f} D={d:.4f} A={a:.4f}")
+
+        over25, under25 = v25.predict_ou25(feature_row)
+        signal["OU25"] = _enrich_signal_entry({"Over": over25, "Under": under25})
+        print(f"  [V25-SIGNAL] OU25 → O={over25:.4f} U={under25:.4f}")
+
+        btts_y, btts_n = v25.predict_btts(feature_row)
+        signal["BTTS"] = _enrich_signal_entry({"Yes": btts_y, "No": btts_n})
+        print(f"  [V25-SIGNAL] BTTS → Y={btts_y:.4f} N={btts_n:.4f}")
+
+        # Additional markets via generic predict_market
+        for model_key, label_map in [
+            ("ou15", {"Over": 0, "Under": None}),
+            ("ou35", {"Over": 0, "Under": None}),
+            ("ht_result", {"1": 0, "X": 1, "2": 2}),
+            ("ht_ou05", {"Over": 0, "Under": None}),
+            ("ht_ou15", {"Over": 0, "Under": None}),
+            ("htft", None),
+            ("cards_ou45", {"Over": 0, "Under": None}),
+            ("handicap_ms", {"1": 0, "X": 1, "2": 2}),
+            ("odd_even", {"Odd": 0, "Even": None}),
+        ]:
+            out_key = self._V25_KEY_MAP.get(model_key, model_key.upper())
+            if not v25.has_market(model_key):
+                continue
+            raw = v25.predict_market(model_key, feature_row)
+            if raw is None:
+                continue
+
+            if label_map is None:
+                # HTFT — 9 combinations
+                htft_labels = ["1/1", "1/X", "1/2", "X/1", "X/X", "X/2", "2/1", "2/X", "2/2"]
+                probs_dict = {}
+                for i, label in enumerate(htft_labels):
+                    probs_dict[label] = float(raw[i]) if i < len(raw) else 0.0
+                signal[out_key] = _enrich_signal_entry(probs_dict)
+            elif len(label_map) == 2:
+                # Binary market
+                labels = list(label_map.keys())
+                p = float(raw[0]) if len(raw) >= 1 else None
+                if p is None:
+                    print(f"  [V25-SIGNAL] {out_key} → EMPTY raw output, skipped")
+                    continue
+                signal[out_key] = _enrich_signal_entry({labels[0]: p, labels[1]: 1.0 - p})
+            elif len(label_map) == 3:
+                # 3-class market
+                labels = list(label_map.keys())
+                probs_dict = {}
+                for i, label in enumerate(labels):
+                    if i >= len(raw):
+                        print(f"  [V25-SIGNAL] {out_key} → insufficient probabilities in raw output")
+                        break
+                    probs_dict[label] = float(raw[i])
+                else:
+                    signal[out_key] = _enrich_signal_entry(probs_dict)
+
+            if out_key in signal:
+                print(f"  [V25-SIGNAL] {out_key} → {signal[out_key]['probs']}")
+
+        print(f"  [V25-SIGNAL] Total markets with real predictions: {len(signal)}")
+        if not signal:
+            raise RuntimeError("V25 model produced ZERO market predictions — cannot continue")
+
+        return signal
 
     @staticmethod
     def _prob_map(signal: Optional[Dict[str, Any]], market: str, defaults: Dict[str, float]) -> Dict[str, float]:
+        """Extract normalised probabilities from signal.
+
+        If the signal contains real model output for this market, use it.
+        If the market is missing from the signal, log a warning and return
+        the defaults as a LAST RESORT (so the pipeline doesn't crash).
+        The defaults are ONLY used for non-core / secondary markets that
+        may not have a trained model yet (e.g. CARDS, HCAP, OE).
+        """
         market_payload = signal.get(market, {}) if isinstance(signal, dict) else {}
         probs = market_payload.get("probs", {}) if isinstance(market_payload, dict) else {}
         if not isinstance(probs, dict) or not probs:
+            print(f"  ⚠️ [PROB_MAP] Market '{market}' NOT found in V25 signal — model output missing")
             return dict(defaults)
         out = {key: float(probs.get(key, value)) for key, value in defaults.items()}
         total = sum(out.values())
         if total <= 0:
+            print(f"  ⚠️ [PROB_MAP] Market '{market}' has zero total probability")
             return dict(defaults)
         return {key: value / total for key, value in out.items()}
 
@@ -730,7 +902,8 @@ class SingleMatchOrchestrator:
             prediction.cards_confidence,
             prediction.handicap_confidence,
         )
-        lineup_penalty = 12.0 if data.lineup_source == "none" else 7.0 if data.lineup_source == "probable_xi" else 0.0
+        lineup_conf = max(0.0, min(1.0, float(getattr(data, "lineup_confidence", 0.0) or 0.0)))
+        lineup_penalty = 12.0 if data.lineup_source == "none" else max(1.5, (1.0 - lineup_conf) * 8.0) if data.lineup_source == "probable_xi" else 0.0
         referee_penalty = 6.0 if not data.referee_name else 0.0
         parity_penalty = 8.0 if abs(ms_edge) < 0.08 else 0.0
         prediction.risk_score = round(min(100.0, max(10.0, 100.0 - max_market_conf + lineup_penalty + referee_penalty + parity_penalty)), 1)
@@ -747,6 +920,8 @@ class SingleMatchOrchestrator:
         prediction.risk_warnings = []
         if data.lineup_source == "probable_xi":
             prediction.risk_warnings.append("lineup_probable_not_confirmed")
+            if lineup_conf < 0.65:
+                prediction.risk_warnings.append("lineup_projection_low_confidence")
         if data.lineup_source == "none":
             prediction.risk_warnings.append("lineup_unavailable")
         if not data.referee_name:
@@ -1142,7 +1317,9 @@ class SingleMatchOrchestrator:
                         if band_val.get("is_value"):
                             boost = min(8.0, boost + 3.0)  # Triple confirmation extra boost
                         prediction.ms_confidence = min(95.0, prediction.ms_confidence + boost)
-                        base_package["prediction"]["ms_confidence"] = prediction.ms_confidence
+                        market_board = base_package.get("market_board")
+                        if isinstance(market_board, dict) and isinstance(market_board.get("MS"), dict):
+                            market_board["MS"]["confidence"] = round(float(prediction.ms_confidence), 1)
                         base_package["v27_engine"]["consensus"] = "AGREE"
                     else:
                         base_package["v27_engine"]["consensus"] = "DISAGREE"
@@ -1157,8 +1334,10 @@ class SingleMatchOrchestrator:
                 base_package.setdefault("analysis_details", {})
                 base_package["analysis_details"]["v27_loaded"] = False
 
-        mode = str(getattr(self, "engine_mode", "v25") or "v25").lower()
-        if mode not in {"v25", "v26", "dual"}:
+        base_package = self._apply_upper_brain_guards(base_package)
+
+        mode = str(getattr(self, "engine_mode", "v28-pro-max") or "v28-pro-max").lower()
+        if mode not in {"v25", "v26", "dual", "v28", "v28-pro-max"}:
             mode = "v25"
 
         quality = base_package.get("data_quality", self._compute_data_quality(data))
@@ -1185,6 +1364,304 @@ class SingleMatchOrchestrator:
             return merged
         return base_package
 
+    def _apply_upper_brain_guards(self, package: Dict[str, Any]) -> Dict[str, Any]:
+        return BettingBrain().judge(package)
+
+        v27_engine = package.get("v27_engine")
+        if not isinstance(v27_engine, dict) or not v27_engine.get("triple_value"):
+            return package
+
+        guarded = dict(package)
+        vetoed_keys = set()
+        guarded_keys = set()
+
+        def mark_guard(item: Dict[str, Any]) -> Dict[str, Any]:
+            if not isinstance(item, dict):
+                return item
+
+            out = dict(item)
+            assessment = self._upper_brain_assessment(out, guarded)
+            if not assessment.get("applies"):
+                return out
+
+            key = f"{out.get('market')}:{out.get('pick')}"
+            guarded_keys.add(key)
+            out["upper_brain"] = assessment
+
+            reason_key = "decision_reasons" if "decision_reasons" in out else "reasons"
+            reasons = list(out.get(reason_key) or [])
+            for reason in assessment.get("reason_codes", []):
+                if reason not in reasons:
+                    reasons.append(reason)
+            out[reason_key] = reasons[:6]
+
+            if assessment.get("veto"):
+                vetoed_keys.add(key)
+                out["playable"] = False
+                out["stake_units"] = 0.0
+                out["bet_grade"] = "PASS"
+                out["is_guaranteed"] = False
+                out["pick_reason"] = "upper_brain_veto"
+                if "signal_tier" in out:
+                    out["signal_tier"] = "PASS"
+            elif assessment.get("downgrade"):
+                out["is_guaranteed"] = False
+                if out.get("signal_tier") == "CORE":
+                    out["signal_tier"] = "LEAN"
+                if out.get("pick_reason") == "high_accuracy_market":
+                    out["pick_reason"] = "upper_brain_downgraded"
+
+            return out
+
+        main_pick = mark_guard(guarded.get("main_pick") or {})
+        value_pick = mark_guard(guarded.get("value_pick") or {}) if guarded.get("value_pick") else None
+        supporting = [
+            mark_guard(row)
+            for row in list(guarded.get("supporting_picks") or [])
+            if isinstance(row, dict)
+        ]
+        bet_summary = [
+            mark_guard(row)
+            for row in list(guarded.get("bet_summary") or [])
+            if isinstance(row, dict)
+        ]
+
+        main_safe = bool(main_pick and main_pick.get("playable") and not main_pick.get("upper_brain", {}).get("veto"))
+        if not main_safe:
+            candidates = [
+                row for row in supporting
+                if row.get("playable")
+                and not row.get("upper_brain", {}).get("veto")
+                and float(row.get("odds", 0.0) or 0.0) >= 1.30
+            ]
+            candidates.sort(key=lambda row: float(row.get("play_score", 0.0) or 0.0), reverse=True)
+            if candidates:
+                main_pick = dict(candidates[0])
+                main_pick["is_guaranteed"] = False
+                main_pick["pick_reason"] = "upper_brain_reselected"
+                reasons = list(main_pick.get("decision_reasons") or [])
+                if "upper_brain_reselected_after_veto" not in reasons:
+                    reasons.append("upper_brain_reselected_after_veto")
+                main_pick["decision_reasons"] = reasons[:6]
+            elif main_pick:
+                main_pick["is_guaranteed"] = False
+                main_pick["pick_reason"] = "upper_brain_no_safe_pick"
+
+        if main_pick:
+            supporting = [
+                row for row in supporting
+                if not (
+                    row.get("market") == main_pick.get("market")
+                    and row.get("pick") == main_pick.get("pick")
+                )
+            ][:6]
+
+        guarded["main_pick"] = main_pick if main_pick else None
+        guarded["value_pick"] = value_pick
+        guarded["supporting_picks"] = supporting
+        guarded["bet_summary"] = bet_summary
+
+        playable = bool(main_pick and main_pick.get("playable") and not main_pick.get("upper_brain", {}).get("veto"))
+        advice = dict(guarded.get("bet_advice") or {})
+        advice["playable"] = playable
+        advice["suggested_stake_units"] = float(main_pick.get("stake_units", 0.0)) if playable else 0.0
+        if playable:
+            advice["reason"] = "playable_pick_found"
+        elif vetoed_keys:
+            advice["reason"] = "upper_brain_no_safe_pick"
+        else:
+            advice["reason"] = "no_bet_conditions_met"
+        guarded["bet_advice"] = advice
+
+        guarded["upper_brain"] = {
+            "applied": True,
+            "guarded_count": len(guarded_keys),
+            "vetoed_count": len(vetoed_keys),
+            "vetoed": sorted(vetoed_keys)[:8],
+            "rules": {
+                "min_band_sample": 8,
+                "max_v25_v27_divergence": 0.18,
+                "dc_requires_triple_value": True,
+            },
+        }
+        guarded.setdefault("analysis_details", {})
+        guarded["analysis_details"]["upper_brain_guards_applied"] = True
+        guarded["analysis_details"]["upper_brain_vetoed_count"] = len(vetoed_keys)
+        return guarded
+
+    def _upper_brain_assessment(
+        self,
+        item: Dict[str, Any],
+        package: Dict[str, Any],
+    ) -> Dict[str, Any]:
+        market = str(item.get("market") or "")
+        pick = str(item.get("pick") or "")
+        if not market or not pick:
+            return {"applies": False}
+
+        v27_engine = package.get("v27_engine") or {}
+        triple_value = v27_engine.get("triple_value") or {}
+        model_prob = self._upper_brain_market_probability(item, package)
+        v27_prob = self._upper_brain_v27_probability(market, pick, v27_engine)
+        triple_key = self._upper_brain_triple_key(market, pick)
+        triple = triple_value.get(triple_key) if triple_key else None
+
+        veto = False
+        downgrade = False
+        reasons: List[str] = []
+        divergence = None
+
+        if model_prob is not None and v27_prob is not None:
+            divergence = abs(float(model_prob) - float(v27_prob))
+            if divergence >= 0.18:
+                veto = True
+                reasons.append("upper_brain_v25_v27_divergence")
+            elif divergence >= 0.12:
+                downgrade = True
+                reasons.append("upper_brain_v25_v27_warning")
+
+        if isinstance(triple, dict):
+            band_sample = int(float(triple.get("band_sample", 0) or 0))
+            is_value = bool(triple.get("is_value"))
+            if market == "DC":
+                if band_sample < 8:
+                    veto = True
+                    reasons.append("upper_brain_band_sample_too_low")
+                elif not is_value:
+                    veto = True
+                    reasons.append("upper_brain_triple_value_rejected")
+            elif market in {"MS", "OU25"} and band_sample > 0 and band_sample < 8:
+                downgrade = True
+                reasons.append("upper_brain_band_sample_thin")
+            elif market in {"OU15", "HT_OU05"} and band_sample < 8:
+                downgrade = True
+                reasons.append("upper_brain_band_sample_thin")
+
+        consensus = str(v27_engine.get("consensus") or "").upper()
+        if consensus == "DISAGREE" and market in {"MS", "DC"} and not veto:
+            downgrade = True
+            reasons.append("upper_brain_consensus_disagree")
+
+        applies = bool(reasons or triple is not None or v27_prob is not None)
+        return {
+            "applies": applies,
+            "veto": veto,
+            "downgrade": downgrade,
+            "reason_codes": reasons,
+            "model_prob": round(float(model_prob), 4) if model_prob is not None else None,
+            "v27_prob": round(float(v27_prob), 4) if v27_prob is not None else None,
+            "divergence": round(float(divergence), 4) if divergence is not None else None,
+            "triple_key": triple_key,
+            "triple_value": triple,
+        }
+
+    def _upper_brain_market_probability(
+        self,
+        item: Dict[str, Any],
+        package: Dict[str, Any],
+    ) -> Optional[float]:
+        raw_prob = item.get("probability")
+        if raw_prob is not None:
+            try:
+                return float(raw_prob)
+            except (TypeError, ValueError):
+                pass
+
+        market = str(item.get("market") or "")
+        pick = str(item.get("pick") or "")
+        board = package.get("market_board") or {}
+        payload = board.get(market) if isinstance(board, dict) else None
+        probs = payload.get("probs") if isinstance(payload, dict) else None
+        if not isinstance(probs, dict):
+            return None
+
+        prob_key = self._upper_brain_prob_key(market, pick)
+        if prob_key is None:
+            return None
+        try:
+            return float(probs.get(prob_key))
+        except (TypeError, ValueError):
+            return None
+
+    def _upper_brain_v27_probability(
+        self,
+        market: str,
+        pick: str,
+        v27_engine: Dict[str, Any],
+    ) -> Optional[float]:
+        predictions = v27_engine.get("predictions") or {}
+        ms = predictions.get("ms") or {}
+        ou25 = predictions.get("ou25") or {}
+
+        if market == "MS":
+            return self._safe_float(ms.get({"1": "home", "X": "draw", "2": "away"}.get(pick, "")))
+        if market == "DC":
+            if pick == "1X":
+                return self._safe_float(ms.get("home"), 0.0) + self._safe_float(ms.get("draw"), 0.0)
+            if pick == "X2":
+                return self._safe_float(ms.get("draw"), 0.0) + self._safe_float(ms.get("away"), 0.0)
+            if pick == "12":
+                return self._safe_float(ms.get("home"), 0.0) + self._safe_float(ms.get("away"), 0.0)
+        if market == "OU25":
+            prob_key = self._upper_brain_prob_key(market, pick)
+            return self._safe_float(ou25.get(prob_key)) if prob_key else None
+        return None
+
+    @staticmethod
+    def _upper_brain_prob_key(market: str, pick: str) -> Optional[str]:
+        pick_norm = str(pick or "").strip().casefold()
+        if market in {"MS", "HT", "HCAP"}:
+            return pick if pick in {"1", "X", "2"} else None
+        if market == "DC":
+            return pick.upper() if pick.upper() in {"1X", "X2", "12"} else None
+        if market in {"OU15", "OU25", "OU35", "HT_OU05", "HT_OU15", "CARDS"}:
+            if "over" in pick_norm or "st" in pick_norm:
+                return "over"
+            if "under" in pick_norm or "alt" in pick_norm:
+                return "under"
+        if market == "BTTS":
+            if "yes" in pick_norm or "var" in pick_norm:
+                return "yes"
+            if "no" in pick_norm or "yok" in pick_norm:
+                return "no"
+        if market == "OE":
+            if "odd" in pick_norm or "tek" in pick_norm:
+                return "odd"
+            if "even" in pick_norm or "ift" in pick_norm:
+                return "even"
+        if market == "HTFT" and "/" in pick:
+            return pick
+        return None
+
+    def _upper_brain_triple_key(self, market: str, pick: str) -> Optional[str]:
+        prob_key = self._upper_brain_prob_key(market, pick)
+        if market == "MS":
+            return {"1": "home", "2": "away"}.get(pick)
+        if market == "DC":
+            return f"dc_{pick.lower()}" if pick.upper() in {"1X", "X2", "12"} else None
+        if market in {"OU15", "OU25", "OU35"} and prob_key == "over":
+            return f"{market.lower()}_over"
+        if market == "BTTS" and prob_key == "yes":
+            return "btts_yes"
+        if market == "HT":
+            return {"1": "ht_home", "2": "ht_away"}.get(pick)
+        if market in {"HT_OU05", "HT_OU15"} and prob_key == "over":
+            return f"{market.lower()}_over"
+        if market == "OE" and prob_key == "odd":
+            return "oe_odd"
+        if market == "CARDS" and prob_key == "over":
+            return "cards_over"
+        if market == "HTFT" and "/" in pick:
+            return f"htft_{pick.replace('/', '').lower()}"
+        return None
+
+    @staticmethod
+    def _safe_float(value: Any, default: Optional[float] = None) -> Optional[float]:
+        try:
+            return float(value)
+        except (TypeError, ValueError):
+            return default
+
     def analyze_match_htms(self, match_id: str) -> Optional[Dict[str, Any]]:
         """
         HT/MS focused response for upset-hunting workflows.
@@ -2104,7 +2581,7 @@ class SingleMatchOrchestrator:
                     return None
 
                 odds_data = self._extract_odds(cur, row)
-                home_lineup, away_lineup, lineup_source = self._extract_lineups(cur, row)
+                home_lineup, away_lineup, lineup_source, lineup_confidence = self._extract_lineups(cur, row)
                 sidelined = self._parse_json_dict(row.get("sidelined"))
                 match_date_ms = int(row.get("match_date_ms") or 0)
                 league_id = str(row.get("league_id")) if row.get("league_id") else None
@@ -2159,6 +2636,7 @@ class SingleMatchOrchestrator:
                     status=str(row.get("status") or ""),
                     state=row.get("state"),
                     substate=row.get("substate"),
+                    lineup_confidence=lineup_confidence,
                     current_score_home=(
                         int(row.get("score_home"))
                         if row.get("score_home") is not None
@@ -2291,48 +2769,78 @@ class SingleMatchOrchestrator:
         self,
         cur: RealDictCursor,
         row: Dict[str, Any],
-    ) -> Tuple[Optional[List[str]], Optional[List[str]], str]:
+    ) -> Tuple[Optional[List[str]], Optional[List[str]], str, float]:
         live_lineups = row.get("lineups")
-        home, away = self._parse_lineups_json(live_lineups)
-        if (home and len(home) >= 9) and (away and len(away) >= 9):
-            return home, away, "confirmed_live"
-
-        # fallback 1: current match participation table
-        cur.execute(
-            """
-            SELECT team_id, player_id
-            FROM match_player_participation
-            WHERE match_id = %s
-              AND is_starting = true
-            """,
-            (row["match_id"],),
+        status_upper = str(row.get("status") or "").upper()
+        state_upper = str(row.get("state") or "").upper()
+        substate_upper = str(row.get("substate") or "").upper()
+        can_trust_feed_lineups = (
+            status_upper in {"LIVE", "1H", "2H", "HT", "FT", "FINISHED"}
+            or state_upper in {"LIVE", "FIRSTHALF", "SECONDHALF", "POSTGAME", "POST_GAME"}
+            or substate_upper in {"LIVE", "FIRSTHALF", "SECONDHALF"}
         )
+        home, away = self._parse_lineups_json(live_lineups) if can_trust_feed_lineups else (None, None)
+        if (home and len(home) >= 9) and (away and len(away) >= 9):
+            return home, away, "confirmed_live", 1.0
+
         home_id = str(row["home_team_id"])
         away_id = str(row["away_team_id"])
-        rows = cur.fetchall()
-        if rows:
-            home_players = [str(r["player_id"]) for r in rows if str(r["team_id"]) == home_id]
-            away_players = [str(r["player_id"]) for r in rows if str(r["team_id"]) == away_id]
-            if not home and home_players:
-                home = home_players
-            if not away and away_players:
-                away = away_players
-            if (home and len(home) >= 9) and (away and len(away) >= 9):
-                return home, away, "confirmed_participation"
+
+        # fallback 1: current match participation table.
+        # Trust this only for live/finished matches; pre-match rows can be stale feed snapshots.
+        if can_trust_feed_lineups:
+            cur.execute(
+                """
+                SELECT team_id, player_id
+                FROM match_player_participation
+                WHERE match_id = %s
+                  AND is_starting = true
+                """,
+                (row["match_id"],),
+            )
+            rows = cur.fetchall()
+            if rows:
+                home_players = [str(r["player_id"]) for r in rows if str(r["team_id"]) == home_id]
+                away_players = [str(r["player_id"]) for r in rows if str(r["team_id"]) == away_id]
+                if not home and home_players:
+                    home = home_players
+                if not away and away_players:
+                    away = away_players
+                if (home and len(home) >= 9) and (away and len(away) >= 9):
+                    return home, away, "confirmed_participation", 0.98
 
         # fallback 2: probable XI from historical starts before match date
         before_date_ms = int(row.get("match_date_ms") or 0)
+        sidelined = self._parse_json_dict(row.get("sidelined")) or {}
+        home_excluded = self._sidelined_player_ids(sidelined.get("homeTeam"))
+        away_excluded = self._sidelined_player_ids(sidelined.get("awayTeam"))
         used_probable = False
-        if not home:
-            home = self._build_probable_xi(cur, home_id, before_date_ms)
+        home_conf = 0.0
+        away_conf = 0.0
+        if not home or len(home) < 9:
+            home, home_conf = self._build_probable_xi(
+                cur,
+                home_id,
+                before_date_ms,
+                excluded_player_ids=home_excluded,
+            )
             used_probable = used_probable or bool(home)
-        if not away:
-            away = self._build_probable_xi(cur, away_id, before_date_ms)
+        if not away or len(away) < 9:
+            away, away_conf = self._build_probable_xi(
+                cur,
+                away_id,
+                before_date_ms,
+                excluded_player_ids=away_excluded,
+            )
             used_probable = used_probable or bool(away)
 
         if used_probable:
-            return home, away, "probable_xi"
-        return home, away, "none"
+            inferred_conf = min(
+                home_conf if home else 0.0,
+                away_conf if away else 0.0,
+            )
+            return home, away, "probable_xi", inferred_conf
+        return home, away, "none", 0.0
 
     def _calculate_team_form(
         self,
@@ -2445,35 +2953,172 @@ class SingleMatchOrchestrator:
         cur: RealDictCursor,
         team_id: str,
         before_date_ms: int,
-        max_days: int = 30,
-    ) -> Optional[List[str]]:
+        match_limit: int = 5,
+        lookback_days: int = 370,
+        max_staleness_days: int = 120,
+        excluded_player_ids: Optional[Set[str]] = None,
+    ) -> Tuple[Optional[List[str]], float]:
         if not team_id:
-            return None
+            return None, 0.0
+        min_date_ms = max(0, before_date_ms - (lookback_days * 24 * 60 * 60 * 1000))
 
-        min_date_ms = max(0, before_date_ms - (max_days * 24 * 60 * 60 * 1000))
         cur.execute(
             """
             SELECT
                 mpp.player_id,
-                COUNT(*) AS starts,
-                MAX(m.mst_utc) AS last_start_ms
+                m.id AS match_id,
+                m.mst_utc,
+                m.home_team_id,
+                m.away_team_id
             FROM match_player_participation mpp
             JOIN matches m ON m.id = mpp.match_id
             WHERE mpp.team_id = %s
               AND mpp.is_starting = true
-              AND m.status = 'FT'
-              AND m.mst_utc < %s
-              AND m.mst_utc >= %s
-            GROUP BY mpp.player_id
-            ORDER BY starts DESC, last_start_ms DESC
-            LIMIT 11
+              AND NOT EXISTS (
+                  SELECT 1
+                  FROM match_player_participation later_mpp
+                  JOIN matches later_m ON later_m.id = later_mpp.match_id
+                  WHERE later_mpp.player_id = mpp.player_id
+                    AND later_mpp.team_id <> %s
+                    AND later_m.mst_utc > m.mst_utc
+                    AND later_m.mst_utc < %s
+                    AND (
+                      later_m.status = 'FT'
+                      OR later_m.state = 'postGame'
+                      OR (later_m.score_home IS NOT NULL AND later_m.score_away IS NOT NULL)
+                    )
+              )
+              AND m.id IN (
+                  SELECT m2.id
+                  FROM matches m2
+                  JOIN match_player_participation recent_mpp
+                    ON recent_mpp.match_id = m2.id
+                   AND recent_mpp.team_id = %s
+                   AND recent_mpp.is_starting = true
+                  WHERE (m2.home_team_id = %s OR m2.away_team_id = %s)
+                    AND (
+                      m2.status = 'FT'
+                      OR m2.state = 'postGame'
+                      OR (m2.score_home IS NOT NULL AND m2.score_away IS NOT NULL)
+                    )
+                    AND m2.mst_utc < %s
+                    AND m2.mst_utc >= %s
+                  GROUP BY m2.id
+                  HAVING COUNT(recent_mpp.*) >= 9
+                  ORDER BY MAX(m2.mst_utc) DESC
+                  LIMIT %s
+              )
+            ORDER BY m.mst_utc DESC
             """,
-            (team_id, before_date_ms, min_date_ms),
+            (
+                team_id,
+                team_id,
+                before_date_ms,
+                team_id,
+                team_id,
+                team_id,
+                before_date_ms,
+                min_date_ms,
+                match_limit,
+            ),
         )
         rows = cur.fetchall()
         if not rows:
-            return None
-        return [str(r["player_id"]) for r in rows]
+            return None, 0.0
+
+        latest_mst = max(int(row.get("mst_utc") or 0) for row in rows)
+        age_days = (before_date_ms - latest_mst) / (24 * 60 * 60 * 1000)
+        stale_projection = age_days > max_staleness_days
+
+        excluded = {str(pid) for pid in (excluded_player_ids or set()) if pid}
+        match_order: Dict[str, int] = {}
+        for row in rows:
+            match_id = str(row["match_id"])
+            if match_id not in match_order:
+                match_order[match_id] = len(match_order)
+
+        player_scores: Dict[str, Dict[str, float]] = {}
+        for row in rows:
+            player_id = str(row["player_id"])
+            if player_id in excluded:
+                continue
+
+            idx = match_order.get(str(row["match_id"]), match_limit)
+            recency_weight = max(1.0, float(match_limit - idx))
+            score = recency_weight
+            if idx == 0:
+                score += 3.0
+            elif idx == 1:
+                score += 1.5
+
+            stats = player_scores.setdefault(
+                player_id,
+                {
+                    "score": 0.0,
+                    "starts": 0.0,
+                    "last_seen_rank": float(idx),
+                },
+            )
+            stats["score"] += score
+            stats["starts"] += 1.0
+            stats["last_seen_rank"] = min(stats["last_seen_rank"], float(idx))
+
+        if not player_scores:
+            return None, 0.0
+
+        ranked = sorted(
+            player_scores.items(),
+            key=lambda item: (
+                item[1]["score"],
+                item[1]["starts"],
+                -item[1]["last_seen_rank"],
+            ),
+            reverse=True,
+        )
+        lineup = [player_id for player_id, _ in ranked[:11]]
+
+        coverage = min(1.0, len(lineup) / 11.0)
+        available_matches = max(1, len(match_order))
+        history_score = min(1.0, available_matches / float(match_limit))
+        core_stability = 0.0
+        if ranked:
+            stable_core = sum(1 for _, stats in ranked[:11] if stats["starts"] >= 2.0)
+            core_stability = stable_core / 11.0
+
+        staleness_factor = max(
+            0.35,
+            min(1.0, float(max_staleness_days) / max(age_days, 1.0)),
+        )
+        confidence = (
+            (coverage * 0.45) + (history_score * 0.25) + (core_stability * 0.30)
+        ) * staleness_factor
+        if excluded:
+            confidence *= 0.92
+
+        confidence_cap = 0.58 if stale_projection else 0.88
+        return lineup or None, round(max(0.0, min(confidence_cap, confidence)), 3)
+
+    @staticmethod
+    def _sidelined_player_ids(team_data: Any) -> Set[str]:
+        if not isinstance(team_data, dict):
+            return set()
+        players = team_data.get("players")
+        if not isinstance(players, list):
+            return set()
+
+        ids: Set[str] = set()
+        for player in players:
+            if not isinstance(player, dict):
+                continue
+            player_id = (
+                player.get("playerId")
+                or player.get("player_id")
+                or player.get("id")
+                or player.get("personId")
+            )
+            if player_id:
+                ids.add(str(player_id))
+        return ids
 
     def _parse_odds_json(self, odds_json: Any) -> Dict[str, float]:
         odds_json = self._parse_json_dict(odds_json)
@@ -4267,7 +4912,8 @@ class SingleMatchOrchestrator:
         lineup_sensitive = market in ("MS", "BTTS", "HT", "HTFT")
         lineup_penalty = 5.0 if lineup_missing and lineup_sensitive else 0.0
         if data.lineup_source == "probable_xi" and lineup_sensitive:
-            lineup_penalty += 4.0
+            lineup_conf = max(0.0, min(1.0, float(getattr(data, "lineup_confidence", 0.0) or 0.0)))
+            lineup_penalty += max(1.0, (1.0 - lineup_conf) * 5.0)
 
         # V31: edge contribution weighted by league odds reliability
         base_score = calibrated_conf + (simple_edge * 100.0 * edge_multiplier)
@@ -4438,8 +5084,11 @@ class SingleMatchOrchestrator:
         away_n = len(data.away_lineup or [])
         lineup_score = min(home_n, away_n) / 11.0 if min(home_n, away_n) > 0 else 0.0
         if data.lineup_source == "probable_xi":
-            lineup_score *= 0.55
+            lineup_conf = max(0.0, min(1.0, float(getattr(data, "lineup_confidence", 0.0) or 0.0)))
+            lineup_score *= max(0.45, min(0.88, lineup_conf))
             flags.append("lineup_probable_not_confirmed")
+            if lineup_conf < 0.65:
+                flags.append("lineup_projection_low_confidence")
         elif data.lineup_source == "none":
             flags.append("lineup_unavailable")
         if lineup_score < 0.7:
@@ -4464,6 +5113,7 @@ class SingleMatchOrchestrator:
             "home_lineup_count": home_n,
             "away_lineup_count": away_n,
             "lineup_source": data.lineup_source,
+            "lineup_confidence": round(float(getattr(data, "lineup_confidence", 0.0) or 0.0), 3),
             "flags": flags,
         }