Is the injury designation the model is served (forecast_delivery.participation_evidence: inj_snapshot + ESPN availability via liveinj.STATUS_MAP) the same object as the one it is fitted on (player_week.injury_report_status, nflverse final report), and where they differ, which one did the outcome agree with?

Owner lane: allocation

Question

Is the injury designation the model is served (forecast_delivery.participation_evidence: inj_snapshot + ESPN availability via liveinj.STATUS_MAP) the same object as the one it is fitted on (player_week.injury_report_status, nflverse final report), and where they differ, which one did the outcome agree with?

Prediction

Frozen in the script docstring before the first run (between the 2026-09-24T0115Z claim and the ~0123Z first run; a guessed 0130Z stamp was corrected): P1 completed 2026 weeks are backfilled, only the target week is empty; P2 at K-90m serve and fit disagree on >=10% of labelled rows, driven by ESPN Questionable persisting where the report cleared and game-day Out; P3 serve-Questionable plays within 10pp of fit-era Questionable. REFUTED (P2) if disagreement <3%.

Finding and verdict

TESTED. H-145's premise refined: (a) 'empty for the live season' is stale for completed weeks (nflverse backfills them) and true only for the target week, where serve is the sole source by construction; (b) 'different vocabulary' is real (79% of labelled rows disagree at K-90m) but costless on 2026 wk1-2 -- all 141 serve-only Out rows sat out; (c) the defect that carries a price is timing coverage: at K-24h 28 of 47 report-designated players (16 Out/Doubtful, 0 played) are served as undesignated. n is two weeks of one season; no held-out season exists because the serve feeds began 2026-09-09.

Reasoning

The hypothesis blamed vocabulary; the cross-tab shows the vocabulary gap is information gain at game time and the loss is a capture-clock gap, which is a different fix (report capture after Friday/Saturday filings, H-139) than a vocabulary map.

Next action

(1) Price the K-24h gap through the model: replay p_play (playermodel.lines_for on the frozen release) for the 28 rows with the report label restored vs served NONE, graded on played -- lands on hurdle_pred's layer, which has no 2026 rows yet. (2) Repeat at weeks 3+ once inj_snapshot is captured after the Friday report; if the Friday capture lands, the K-24h gap should close, which is the falsifier for the timing mechanism. (3) The 2 NONE->Questionable rows are the only sticky-status evidence; watch it under the 5-per-team cap.