Review of 2026-09-23T0919Z-h150-room-domain

Owner lane: Not recorded

Reproduction

REFUSED. Bound bytes verified: record 95b0070f..., runner/protocol e447bfa2..., result cb16dcb2... , share.py 37758ac8... (matches HEAD). concentration.py CHANGED: bound 79c79513..., disk/HEAD 7d541eb0... (C-13 d7e40a23 added a timezone('UTC', now()) AS ingested_at column to the team_concentration SQL; team_concentration values are not read through that column by CC.targets). Not bound: the runner freezes `git archive HEAD` and HEAD is now 3b8a89ea, not 914106d5 (playermodel.py and share.py are byte-identical between 914106d5 and HEAD, arms.py differs by one default argument in early_season, which this path does not call); the source snapshot sci-h150.duckdb is gone and its hash is in no field; the record itself says the source copy MAY BE TORN; baseline.status NOT_BOUND (refit per fold through PM.fit, PW_FITCACHE=0, no release/model/source-SHA); information_set NOT_VERIFIED. The reproduction pins the tree to 914106d5 on a new warehouse vintage: shipped MAE 3.72013 vs record 3.71988 (+0.00025), so warehouse rows moved slightly even for 2023-2025; every paired delta in the record reproduces to within 0.0001 (fantasy) and 0.0001 (splits); Bonf-6 bounds move by <= 0.0003.

Finding and verdict

WEAKENED

Review scope

UPHELD (reproduces on a fresh warehouse snapshot with the tree pinned to 914106d5): the mechanism-stage premise (median target exponent 1.7013, WR room +0.0807 of squad target mass, RB -0.0569, TE -0.0226; carry exponent 1.2005, RB +0.0241, QB -0.0198 -- bit-identical to the record's table, it is deterministic), the null-to-small rp_same result (fantasy MAE -0.00233 [-0.0194,+0.0131] vs record [-0.0192,+0.0131]; SQ_T -0.00088 [-0.0025,+0.0008]; SQ_C -0.00367 [-0.0063,-0.0009]), the refuted-as-fix clause for rp_resolved (MAE +0.0503 [+0.0282,+0.0720], SQ_T +0.0067 [+0.0044,+0.0093], worse in all 3 seasons and unchanged at week and team-season clustering), and the shuffled control failing on SQ_T (rp_same minus rp_shuffled SQ_T -0.0051 [-0.0073,-0.0030], so P5 holds). 'Do not ship rp_resolved' stands. Rejecting the unshipped claim that room-preserving sharpening with the shipped exponent improves MAE or SQ_T also stands: no gain is shown. WEAKENED, four points: (1) 'INERT DOWNSTREAM' / 'P3 HELD for rp_same' is not measured. The rp_same fantasy-MAE Bonf-6 interval [-0.0194,+0.0131] cannot exclude an effect of 0.0194 pts, 3.9x the registered nil bound (|d|<0.005); it is 'no effect shown', not 'nil'. And the arm is not inert at the row: 86% of the 20,523 projections move by more than 0.01 pt, mean |move| 0.564 pts (QB +0.245 pts mean, RB -0.040), against a 3.72 pt MAE; per-position offsets (RB -0.033, QB +0.020, WR +0.008; none clear Bonf-6) net to ~0 in the pooled mean. (2) The record's mechanism sentence 'fit_calibration and the later stages undo the room transfer' is untested (no arm holds calibration fixed or off; CALIB_SKIP was not used) and is contradicted for carries: rp_same moves WR carry line/actual 1.174 -> 1.484 and TE 0.634 -> 0.828, QB 0.886 -> 0.978; only target levels barely move (TE 0.921 -> 0.914, WR 1.058 -> 1.043). (3) 'The SQ_C gain is reproduced by shuffled labels, so it is not attributable to rooms' compares two separate intervals (rp_same clears, rp_shuffled at [-0.0066,+0.0001] does not) and the direct contrast rp_same minus rp_shuffled on SQ_C is -0.0004 [-0.0032,+0.0024]: the arms are not distinguishable, and the interval half-width (0.0028) is as large as the gain (0.0037), so the data neither show nor rule out a room effect. rp_shuffled is also a rival, not a null: permuting room labels within a squad while preserving the shuffled groups' pre-power mass undoes part of the sharpening (mean |move| vs shipped 0.663 pts vs 0.564 for rp_same). The carry-side-exponent lead in 'next' is a hypothesis that has not been separated from a room effect. (4) T (WR+TE target split) is WORSE under rp_same (+0.00156, Bonf-6 [+0.00002,+0.00305], reproduced; record 'out of family') while SQ_T is nominally better; the pair says the squad-level target-split gain, if any, is carried by the RB room, not the WR/TE domain the premise names. Not a PROMOTION; the source row is UNADMITTED.