ACCOUNTABILITY
MODEL VERDICTS
Every model here is judged VALIDATED, PROVISIONAL, DEPRECATED, or REJECTED — and every verdict links the committed evidence behind it.
This is the model-level accountability layer. The call-level layer is The Ledger and Call-Up Receipts — every individual call, tracked to an outcome. Together they are two layers of accountability, not one.
The published failures are the point, not a footnote: a model that is losing or untrainable gets a first-class row, same as any win. See how the verdicts work. Registry updated AUGUST 14, 2026.
| Model | Verdict | Why | Evidence | As of |
|---|---|---|---|---|
| MLB Projection (H+P / Marcel) feeds value | PROVISIONAL | Beats internal baselines on held-out data but has not yet beaten Steamer; the live forward check is running and currently losing, below the maturation gate. | Forward-gate source comparison | SEPTEMBER 13, 2026 |
| Prospect Rank v1 feeds value | PROVISIONAL | Calibration is under review and coverage is still candidate-stage; the board ships, the proof is not yet in. | Rank v1 calibration report | SEPTEMBER 13, 2026 |
| Peak Projection v1 | PROVISIONAL | Display-only role and skill-shape context under active calibration review. It does not feed rank or value and is not a validated stat forecast. | Peak Projection calibration report | SEPTEMBER 13, 2026 |
| Universal Prospect Model — Hitters feeds value | VALIDATED | Stage 1 ordering resolved on 3,652 matured 2016–2021 outcomes: beats the level/age prior AND the 25-historical-neighbors baseline on Spearman, Kendall tau-b, and ROC AUC, every delta interval excluding zero (player-clustered bootstrap, n=1,765 hitters, 5 closed cohorts). Support re-confirmed by the registered 2021-maturation re-run (docs/registration-2026-08-14-stage1-maturation-rerun.md, monitoring clean). Retrospective ordering evidence; not a probability calibration and not a public superiority claim. | Stage 1 outcome proof (hitters) | AUGUST 14, 2026 |
| Universal Prospect Model — Pitchers feeds value | REJECTED | Beyond-neighbors claim resolved negative by the registered maturation re-run (docs/registration-2026-08-14-stage1-maturation-rerun.md, rule 3, executed 2026-08-14): on 1,887 matured outcomes over five closed cohorts (2016–2019, 2021), all three ordering deltas vs the 25-historical-neighbors baseline straddle zero with negative point estimates (Spearman −0.008 [−0.041, +0.024]; Kendall tau-b −0.010 [−0.038, +0.016]; ROC AUC −0.006 [−0.031, +0.019]). The model DOES beat the level/age prior with every interval excluding zero — that weaker claim stands, and served scores are unchanged. Any future beyond-neighbors claim requires a materially changed pitcher model and a new registration. | Stage 1 outcome proof (pitchers: beats level/age prior; indistinguishable from neighbors — claim rejected) | AUGUST 14, 2026 |
| MiLB Translation | VALIDATED | A deterministic, reproducible level-translation port — measured, observe-only, and never a value input. | Raw-data independence audit | SEPTEMBER 13, 2026 |
| Shape Comps | VALIDATED | Deterministic and reproducible — the closest real MLB shapes on three translated rates, measured not modeled. | Shape comps artifact | SEPTEMBER 13, 2026 |
| Pitch Discipline — exact metrics | VALIDATED | Swing/Whiff/SwStr counted directly from play-by-play, ProspectSavant-matched — observe-only, never a value input. | Pitch-discipline layer artifact | SEPTEMBER 13, 2026 |
| Pitch Discipline — estimated zone metrics | PROVISIONAL | Zone rates from a pixel-coordinate calibration, tagged "est." — held-out agreement clears the bar but they are estimates, not measurements. | Pitch-discipline calibration metrics | SEPTEMBER 13, 2026 |
| Ahead-of-the-Curve divergence calls | PROVISIONAL | A live, pre-registered call ledger with the sample still maturing toward its published horizon — open the ledger for the current number. | The Ledger (AOTC scorecard) | SEPTEMBER 13, 2026 |
| Wins (W) target | REJECTED | There is no W in the label seasons, so a W-inclusive preset hits the partial-coverage refusal — ValuCast cannot train a W target and does not fake one. | League-adapter coverage | SEPTEMBER 13, 2026 |
No failure number is restated here — every verdict links the live artifact where the current figure renders. A verdict never outlives its evidence: the daily build fails if any cited artifact goes missing. Back to how ValuCast works.