- Add LeagueTier DB model and Prisma schema
- Add league-tiers service (CRUD, sync, retrain trigger)
- Add league-tiers controller with admin API endpoints
- Add /v1/admin/retrain endpoint in AI engine (extract→train→reload pipeline)
- Retrain V25 Pro with 48 quality leagues (MS accuracy: 26.9%→51.4%)
- Update qualified_leagues.json (443→48 leagues)
- Include V25 model files in repo for Docker deployment
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- OllamaClient utility for llama3.2:3b API calls (timeout 30s, non-fatal)
- OllamaCommentary service builds structured Turkish prompt from prediction data
- PredictionsService enriches response with ai_expert_commentary field
- Frontend prediction-card displays AI commentary section above match_commentary
- goals_form now uses avg of last 5 historical matches instead of current match goals
- squad_quality removes current match goals/assists, uses only pre-match known data
- adds temporal filtering via match_id -> mst_utc mapping
- Load feature columns dynamically from feature_cols.json
- Add 20 missing odds_*_present boolean flags to fallback list
- Fixes LightGBM 'features in data (82) != training data (102)' crash
- Reset consecutiveFailures on cooldown expiry (half-open state)
so a single retry failure doesn't immediately re-open the circuit
- Exclude AI Engine app-level 500s from circuit breaker count
(only network/infra errors: timeout, 502, 503, 504, 429)
- Return null gracefully instead of throwing 503 when no cache exists
- Add DB fallback for non-cooldown AI Engine failures
- Remove blocking wait-and-retry that held requests for up to 20s
- Add 4-level fallback when AI circuit breaker fires cooldown:
1) In-memory cache (10min TTL)
2) DB stored prediction (no TTL filter)
3) DB cached prediction (with model version check)
4) Wait out cooldown + retry once (max 20s wait)
- Raise circuit breaker threshold from 3 to 5 consecutive failures
- Reduce cooldown duration from 30s to 15s for faster recovery
- Add extractCooldownMs helper to parse remaining ms from error detail
- extract_training_data.py: switch from top_leagues.json (23) to qualified_leagues.json (265)
- update_implied_odds.py: new script to backfill implied odds from real market data
- train_score_model.py: rewrite with v25 102-feature set + temporal split
- single_match_orchestrator.py: integrate ML score model with heuristic fallback
Live odds have missing selections (e.g. '1' key removed from Maç Sonucu
after kickoff), causing the AI model to produce wildly incorrect predictions
(e.g. 3.5% home win for Bristol City). Two guards added:
1. fetchOddsForMatches: Exclude live/finished matches from odds fetch query
2. processMatchOdds: Skip odds/lineups/sidelined overwrite if match already
has pre-match odds and is live/finished
fahricansecer