Review of 2026-09-23T1604Z-h076-served-pace-block

Owner lane: Not recorded

Reproduction

REFUSED. The source record has input_artifacts [] and no registration_receipt or attempts; its as-run snapshot fc-h076.duckdb is no longer in pw-snapshots; and two hashed sources moved after f4bca4d2: playerweek/arms.py e782a10b vs recorded d654b285 (diff: one default argument of early_season, unrelated to rows_for/score_arm) and playerweek/grains.sql 9e1b9fc4 vs recorded ffa257d9 (C-13 adds ingested_at columns; tw_lags unchanged). playermodel.py eeeb6a3c and share.py 37758ac8 are byte-identical to the recorded hashes. Bound and unchanged: prereg ee2dceb4, source runner 6c50b435, source out.json 49cb376f, source record 4981e9f7.

Finding and verdict

UPHELD

Review scope

Narrow. Upheld: every number in the record reproduces to the printed digit on a fresh snapshot at HEAD (b-a +0.0037 [-0.0057,+0.0131], c-a +0.0013, d-a +0.0077, X-a +0.0067, team pass/rush contrasts, per-position, per-phase, per-season); arm b is leakage-free (poisoning every team_week row at or after week w changes served() on 0 of 1728 (season, week, team) keys, while the same poison changes the hindsight arm); the planted control is not vacuous (PL-a +0.0168 [+0.0108,+0.0230], the record reported the point only); the frozen falsifier (b-a outside +/-0.015) is not tripped. WEAKENED inside the record, not enough to change the verdict: (1) arm b is not the served path for movers, because latest_rows joins the pace CTE on the player's OLD team and overrides team afterwards; the production-faithful arm bm (wk1-3 movers get the prior-season team's pace) gives bm-a +0.0050 [-0.0046,+0.0144], still nil, but the 702-row mover cell is +0.0267 [-0.0232,+0.0825] and cannot say; (2) 'information present at the volume layer is destroyed before the player line' is unidentified: the player-level 98.33% interval on c-a half-width is 0.0095, larger than the attenuated gain the record claims, so 'destroyed' and 'present but below detection' are not separated; the record's own next step (a perfect team-attempts override) is the intervention that would separate them.

Next action

(1) Run the record's own next step, a perfect team_pass_att_pred override through volume_for, and report the player interval beside it; that is the intervention that separates 'destroyed' from 'below detection'. (2) A mover-only board test for the wk1-3 old-team pace needs more than 702 rows; H-195 (2026-09-26T0624Z) already found movers' targets nil at the served board, so the pace quirk has no visible player-level cost.