The public, graded record.

Live record

9.5% calibration error, 0.207 Brier, 252 grades

Engine change, 2026-07-18 → 2026-07-19. For one day the strikeout model carried an extra feature — the trailing strikeout rate of the opposing starting nine. An adversarial review found it was trained on the lineup each pitcher actually faced but served a posted-or-projected one, and that on the projected path it made the model worse than having no such feature at all; its apparent gain was never distinguishable from zero. It was removed on 2026-07-19 and the model retrained without it. 81 graded line-results (27 starts, all from 7/18) plus the 7/19 board were produced by that version. Those predictions stand exactly as published and stay in the running totals — they were made in good faith before first pitch, and deleting inconvenient history is the one thing this record will not do. Method: docs/K-FEATURE-VERDICT.md.

Backtest, not the live record

2.2% pooled calibration error, per-line 2.2% / 3.9% / 2.6%, 0.205 Brier, 1090 holdout starts

Trained on real starts before 2026-06-01, openers excluded. The three lines per start are nested views of one outcome, so the pooled number can flatter. It earns nothing until the live record proves it.

Read it straight: the persistence baseline is ahead of the model on full-distribution log-loss. Stated precisely, because the earlier wording overstated it — the gap is +0.0169 with a 95% paired bootstrap interval of [-0.0073, +0.0410], so persistence leads on the point estimate but the difference is not distinguishable from zero at this sample size. The model does win the P(over) Briers at every published line. The live record is the referee.

Live record

2.1% calibration error, 0.198 Brier, 1282 grades

Backtest, not the live record

0.5% pooled calibration error, per-line 0.6% / 0.4%, AUC 0.56 / 0.57, 9946 holdout batter-games

Calibrated but only slightly sharper than the base rate. Projected lineups miss the actual nine about 24% of the time; that scratch rate is declared here in advance. Judge the live record against the live-gradeable subset: log-loss 1.2159, ECE 0.6%. Night-one predictions of 7/17 predate that evening's engine fixes; immutable either way.

On this market the model beats both cheap baselines on every metric. On the K market it does not, and both statements are published with the same straight face.

Live record

2.2% calibration error, 0.101 Brier, 641 grades

Backtest, not the live record

0.6% calibration error, 0.590 AUC, 10462 holdout batter-games

Aggregate rate bias -0.3%: predicted 12.7% vs observed 13.0%. The mid-summer power surge outran the training window, reported straight. Live-gradeable subset (76%): log-loss 0.4268, ECE 0.4%.

What the weather features bought (goal HR-2, published 2026-07-19). Split result, reported straight: weather did not measurably improve discrimination — log-loss moved -0.0006 and Brier -0.0001, both smaller than this engine's own run-to-run variation, so they are not evidence of added skill. What it fixed was the level: the summer under-prediction shrank from -0.85pt to -0.31pt, absorbing about 64% of the seasonal bias, and calibration error fell 1.08% → 0.61%. Mostly air density — temperature, humidity and elevation combined into the quantity that actually governs carry. Caveat: the backtest trains on reanalysis (what the weather was) while live serving uses forecasts (what it was predicted to be), and day-ahead wind direction is wrong often enough to flip blowing-out/blowing-in in roughly one game in five. The bias fix rides on density, which forecasts well.

Live record

Accumulating: 0 of 60 grades before the live curve draws.

Scored by CRPS over the whole distribution; the curve here is the 8.5 line only. What clearing the launch gate does and does not mean is stated on the Totals board.

Backtest, not the live record

-0.0212 vs climatology, 95% CI [-0.0413, -0.0007]

Clears the pre-registered floor on 100% of 100 refits. Seed stability is refit reproducibility, not a confidence statement — the interval above is the confidence statement, and it crosses zero.

Mean total 9.06 predicted vs 9.35 actual (-0.29 runs/game, -1.5 sigma): the model tracks the training run environment and the holdout ran hotter. Disclosed because a calibration product that hides a level miss is not a calibration product.

Backtest weather is reanalysis (what actually happened); live serving uses the forecast at publish time, so live lift will be smaller and that gap is not yet measured.