Review of 2026-09-23T1105Z-h223-swap-vs-discount

Owner lane: Not recorded

Reproduction

REFUSED. Record sha e77dee5d... verified; runner bddca728..., protocol df305f7f..., result out.json fb1dc2b7... all verified byte-for-byte on disk. Not bound / changed: (1) input_artifacts is [] -- the source snapshot allocation_h223.duckdb (1598 MB) is gone and its hash is nowhere; the record itself says SNAP warned the copy MAY BE TORN. (2) playerweek/playprob.py bytes changed since binding, 3d838d69 -> de83ca65 (b6ba003d, PW-148, adds a roster= argument, default 'act', so the served path is unchanged; playermodel.py eeeb6a3c and decision_stats.py 695461d2 unchanged). (3) the re-execution used a different warehouse copy (referee_h223.duckdb sha256 5b60e4ec3935c32947a477588358f10aa4570d042a058275cd5cc462efe1cbf5, 2076 MB) and the numbers moved slightly: close pairs 221,522 vs 221,835, eligible 31,064 vs 31,027, room census identical row for row (8601/8577/8688; INA 1093/1069/1174; repeat max abs 0.0). The 313-pair shift is served points crossing the 6.0 floor between warehouse vintages; I did not locate which table changed, and the record's mitigation (census equals H-091's) cannot see it.

Finding and verdict

WEAKENED

Review scope

UPHELD, as a re-execution on a fresh snapshot with the tree pinned to bdd04f82: the FROZEN composite SWAP (order by c = served/p_play, then swap if the pick is INA) is worse than the shipped DISCOUNT on eligible pairs. SWAP-D regret +0.17032 [+0.01965,+0.30662] Bonf-2 (record +0.16203 [+0.01419,+0.29869]), accuracy -0.00782 [-0.01703,+0.00243] spans 0, pooled +0.02641 [+0.00649,+0.04373], pick-differs 1.77% (<3% NIL threshold), QB +0.514, RB +0.149 spans 0 at 95%, WR +0.026 spans 0, TE -0.121 spans 0. PERM control fails as the record says (+0.628 [+0.427,+0.819]). REFUTED, as read: the record's headline reading that the post-inactives SWAP step costs regret ('a post-list swap costs +0.162 regret', 'the served p_play x line is the better start/sit number') is not what the frozen arm measured. SWAP fires on 0.64% of eligible decisions; the cost is the reorder by c. The arm the record lists NOT RUN, a swap on DISCOUNT's own order (SWAPD), gains: SWAPD-D regret -0.02356 [-0.04645,-0.00857], accuracy +0.13pp [+0.05,+0.27] on eligible pairs, -0.00372 [-0.00746,-0.00130] pooled; reorder alone (COND-D, no swap) +0.36544 [+0.17340,+0.55626]; SWAP-COND -0.19512 [-0.30072,-0.09189], so the swap step helps by 0.20 on top of the reorder it was bundled with. On pairs that hold an INA man once the floor is put on c (the record's NOT-RUN arm, 5,307 pairs) DISCOUNT regret is 0.876 vs SWAPD 0.307 (-0.569 [-0.991,-0.394]); D picks the INA man in 15.2% of them. WEAKENED, not REFUTED, because the shipped-vs-frozen-composite comparison stands and the record's own mechanism paragraph names the reorder as the cost; the verdict/why text and commit title over-read it. Separately the CONSENSUS arm is not a decision-time number (see attack 2).