Edges are discovered by sweeping the model's win/place probabilities across price (Betfair BSP), side (back/lay) and market cells (track, grade, trap, distance, day, field size, probability bands), then settled exactly as they'd be bet. An edge qualifies on economics and cross-period consistency — a positive race-clustered bootstrap lower bound on return, profitable across three independent windows (training, early holdout, late holdout) — rather than a multiple-testing gate that was hiding real edges. Every qualifying edge goes to paper trading, tiered by conviction (T1 full stake, T2 observation stake) so live results show where the model truly works. Staking is chosen per edge by a shootout over flat and fractional-Kelly methods, scored on return minus drawdown.
| Surface | Verdict | Evidence |
|---|
Ranked by holdout lower-bound ROI. "Recovered" = an edge the old multiple-testing gate rejected.
| Edge | Bet | n | Holdout ROI | CI-lo | Steady | Cover | Stake |
|---|