project_wnba_season_endgame

Verified 2026-08-06 against SDS /wnba/games/date/: 39 games over 14 slate days, then ZERO from 2026-08-20 onward. ~967 more scoreable props for the entire rest of the regular season (measured 24.8 scoped props/game, 74.4/slate-day).

This makes hit rate useless as a referee. Power at baseline 51.6%, alpha .05, 80%: +2pp needs 4,894 props (66 slate days); +3pp needs 2,173 (29 days); paired McNemar +2pp still needs 2,928 (39 days). All exceed the 14 days left. Any model change made now CANNOT be validated on accuracy this season.

What DOES fit the budget:

• [internal detail removed]
• Paired MAE: per-prop abs-error diff sd = 0.6263 → 0.10 pts needs 308 props (4 days); minimum detectable with the whole remaining season is 0.056 pts.
• NOT Brier: per-prop Brier-diff sd 0.1067 → a 0.005 move needs 3,574 props (48 days).

Brier-vs-market negative REPRODUCED on this window (n=670, 2026-07-23..08-05, run read-only with --no-history): market Brier 0.24807, model raw 0.26176 (−5.52%, paired z=−3.28), model cal 0.25195 (−1.56%, z=−2.20). Skill capture −419%. Model AUC 0.5179 vs market 0.5514 — still cannot rank. Confirms project_brier_vs_vegas_closed.

The shipped calibration map is a CONSTANT. scripts/ops/forecast_calibration.json platt a=−0.0854 b=0.0349 — slope ~0 collapses every forecast to sigmoid(−0.0854)=0.4787; "model cal" resolution is exactly 0.00000 and brier_board.py prints its own shrinkage-only warning. Must not be used to pick or size bets.

Scout accuracy does not clear breakeven: rolling-30d hitRate 0.516 n=4,436 → 95% CI [50.1%, 53.1%], vs −110 breakeven 52.38%. Only z=2.1 above a coin flip. The endpoint has no per-stat/confidence/side breakdowns (only league and date params).

Consequence for planning: ship only harm-removal to WNBA now (see project_wnba_injury_join); do NOT deploy model/semantics changes into the last 13 days — see the deploy-risk note in project_wnba_stranded_deploy. Real model work belongs in the NBA offseason where there is time to measure.