Append-only forecast record · 72 records · calibration scoring · market verdict pending
Champion Brier 0.5274 over n=72 settled group matches. A small sample is suggestive, not a track record.
The group stage is complete. This graded ledger covers only the 72 pre-registered group matches and stays at that count by design: knockout pairings were unknown when the model was locked, so per-match knockout forecasts are published as an explicitly ungraded live surface on the bracket. The pre-registered progression checkpoint (which teams the frozen model favored to advance) will be reported here when it settles.
72 graded forecasts scored · 104 matches settled · snapshot 2026-07-30T01:17Z
Proper Scoring Rules: lower is better
| Metric | M0 | M1 | M2 | M3 | M★ |
|---|---|---|---|---|---|
| Brier | . | . | . | . | 0.5274 |
| Log-loss | . | . | . | . | 0.8993 |
| RPS | . | . | . | . | 0.1596 |
Calibration: Reliability Diagram
Points on the diagonal indicate perfect calibration. Each point represents one decile bin; size reflects sample count. Only bins with n ≥ 1 are rendered.
Default sort: reverse-chronological. HIT / MISS / NEUTRAL rows share identical grid, padding, and type size. §7.2 invariant. No “win rate” or ROI sorting is exposed. Glossary
Pre-registered §7.6: the three forecasts where M★ assigned the highest probability to a modal outcome that did not occur. This is the inverse of a highlight reel: the strongest statement of calibration discipline the ledger makes.