Review of 2026-09-23T1636Z-h067-coach-zero-price

Owner lane: Not recorded

Reproduction

REFUSED. The source record has input_artifacts [] and no registration_receipt, author_id or attempts; the as-run snapshot sci-h067.duckdb is gone from pw-snapshots; playerweek/arms.py is e782a10b now vs d654b285 bound (early_season default argument, inert for rows_for/score_arm); statline.py moved (144 lines) after 17fe17b7; rel_coach was rewritten (ingested_at 2026-09-23T17:19Z, after the 16:41Z run); shipped MAE differs in the 4th decimal (3.7204 vs 3.7201, RB 4.0258 vs 4.0244) with NO playerweek/ code change between e26341a8 and HEAD, so the drift is data. Bound and unchanged: prereg 4a159f30 (mtime 10:37:20 MT, 12s after the runner started and before it finished; outcome blindness is the record's own attestation), runner 39e29596, source out.json c20a6421, source record 8fff7b5d, playermodel.py eeeb6a3c and share.py 37758ac8 byte-identical to the recorded hashes.

Finding and verdict

WEAKENED

Review scope

Headline survives, three secondary claims do not. SURVIVES (non-exact reproduction at HEAD d8c149dc on a fresh snapshot): zero - shipped pooled +0.0137 [99% +0.0023, +0.0257] (record +0.0156 [+0.0038, +0.0278]), interval excludes zero, point above the frozen +0.005 refutation line; RB +0.0303 [+0.0046, +0.0578] is again the only position that excludes zero; replica max|diff| 0.0; n 20,523 and lag coverage 5765/5349/5394 identical to the record. DO NOT SURVIVE: (1) 'carry-forward recovers all but +0.003': the record never tested zero - lag, and the interval it cites (lag - shipped spanning zero) is not an equivalence. Reproduced lag - shipped +0.0045 [-0.0025, +0.0117] (upper edge is 85% of the zero cost); zero - lag +0.0092 [-0.0016, +0.0211] at 99% and [+0.0009, +0.0182] at 95%; lag's recovered share of the zero cost is 67% with a 95% bootstrap interval [14%, 108%]. 'Cheapest writer is a carry-forward' is a recommendation the data leave open, not a measured recovery. (2) 'about 40% of the zero arm's cost is what the coach information is worth, the rest is out-of-distribution 0': not reproduced (shuffle - shipped +0.0023 [-0.0076, +0.0120] is 17% of the zero cost here vs 40% in the record, both intervals span zero), and a serve-time swap of a feature's values does not price the information in the feature in any case (that needs a refit without it). (3) 'behaves like a regime label for an unmatched/novel sideline' is asserted with no intervention; the record's own signed-move prediction failed and the replacement story was not tested against the rival that trees clamp an out-of-range 0 to the lowest bin.

Next action

Before any carry-forward is built: (a) test zero - lag on the first post-flip board, or on more seasons, with an interval that can separate them (this sample's 99% half-width is ~0.012 against a 0.009 gap); (b) run the record's NOT-RUN arm, a refit without coach_pos_* on the same walk-forward, which is the only design that prices the coach information; (c) bind how rel_coach(c,S,w) is built and that it excludes week w.