Review of 2026-09-23T1412Z-h213-served-opp-block

Owner lane: Not recorded

Reproduction

REFUSED as exact: source record input_artifacts is [] (the as-run snapshot bj-h213.duckdb is not bound and is not the object read now: fresh snapshot is 2080 MB vs 1646 MB as-run), and playerweek/arms.py differs from the recorded baseline (record d654b285..., now e782a10b...; commit 5bb12400 changed one default argument in arms.early_season, unrelated to rows_for/score_arm; the runner does not call it). HEAD moved c8a176a8 -> 22efec24 (976 commits). playermodel eeeb6a3c and share 37758ac8 match the record. Protocol sha 2f4028c3, runner sha 487c7678, result sha 25dfbb92 (the run/*.out.json), source record sha a262cb88 all match the packet.

Finding and verdict

WEAKENED

Review scope

UPHELD and REPRODUCED: the frozen prediction (b-a within +/-0.01, REFUTED only if b-a Bonferroni-2 interval clear of +0.01) survives the attempt; every field of the record's result reproduces exactly (b-a -0.0034 [-0.0104,+0.0037], c-a +0.0000 [-0.0061,+0.0061], X-a -0.00003 [-0.0076,+0.0072]; replica max|diff| 0.0 all seasons), and the same nil holds on a squared-error endpoint (below). WEAKENED: (a) the record's mechanism sentence and 'why' -- the trained OPP block 'carries nothing measurable', 'inert', 'no substitution changes accuracy' -- are stronger than the instrument supports: its own planted control (+1 sd on opp_points_allowed_l3, lines move 0.139 pts/row) costs +0.00234 MAE with a paired game-bootstrap 97.5% interval [-0.00233,+0.00684], i.e. the instrument cannot see a planted shift that the model demonstrably reads, so the 'PL non-vacuous' control passed on line movement only and does not establish accuracy resolution; the resolution (half-width ~0.007 MAE, ~0.1 MSE of 32.49) bounds the claim to 'no substitution differs from the real block by more than ~0.007 MAE', which does not exclude a real but small opponent effect (the record's own RIVAL clause); (b) the priority conclusion 'refilling the served opponent block is not worth a build item' is licensed only as 'not worth it at ~0.01 pts/wk MAE', not as 'the block is inert'; (c) the served path differs from arm b in more than the OPP columns (share.feature_sql(team=False) also omits TEAM_FEATURES, TREND_FEATURES, exp_pass_att/exp_rush_att; and latest_rows serves ONE row per man at his last involved week with a whole-season lg mean, a season-ahead projection, not a within-season week), so b prices the OPP substitution only, as the record's title says, not the served row.

Next action

If the served-opponent question is revisited: bind the snapshot as an input_artifact; report a planted-shift interval on the gated endpoint and size the planted shift to the smallest effect that would matter, so the nil carries a detection floor; price the full served row (team=False columns) rather than OPP alone.