Baseball · Models
Historical evidence archive
Training diagnostics belong to the named artifact and cohort. They do not measure the current whole-system release.
Model Trust
Model release and diagnostics
Release evidence for served artifact retrosheet-game-date-bounded-fielding-2026-v16.0.0, led by relative proper scores and reproducibility. Closing-line results remain diagnostic; they do not authorize real-money betting.
Operator-directed reproducibility release after the ordinary relative gate failed: moneyline Brier +0.0005, log loss +0.0010; totals Brier +0.0000, log loss +0.0000. Real-money use is not authorized.
Promotion decision: release-v16-and-update-serving on 2026-08-29. Operator-directed paper release of the byte-reproduced model built from corrected date-bounded fielding history. The ordinary relative proper-score gate remains failed and real-money use remains unauthorized.
Artifact binding: historical-game-symmetric-date-bounded-fielding-v16, sha256:1be510e2c9ea60d52e4054ac88609b7c3aaf903f54d0d6d748d2eb410dcd584a.
Close-market diagnostic: trails historical-closing-totals on 11196/11196 games (model Brier 0.2503 vs market 0.2496, log loss 0.6937 vs 0.6924).
Benchmark evidence: Evidence path recorded; hosted viewer unavailable for this artifact. Source: offline/retrosheet-parser/warehouse/releases/v16/candidate-a5/v16-candidate-evaluation.json · SHA-256 sha256:7b8fdf…c5569f59.
| Market | Games | Market Brier | Model Brier | ROI edge>0 | ROI edge>2% | ROI edge>5% | Verdict |
|---|---|---|---|---|---|---|---|
| Totals (O/U) | 11196 | 0.2496 | 0.2503 | - | - | - | Model is less accurate than the closing line (higher Brier); positive backtest ROI here is likely variance, not edge. |