Review of 2026-09-23T0452Z-h224-roster-window

Owner lane: Not recorded

Reproduction

REFUSED. Source bytes bind: record c7d60ec6..., protocol 9c3b1b97..., runner 18705c0c..., result 5192932f..., lags.sql 499592a9..., h224_lib 35bf7936... all verified unchanged on disk. What is not bound: (1) the input warehouse, the source snapshot sci-h224.duckdb is gone (only a .wal and a -torn copy remain) and its hash is in no field; (2) baseline.status NOT_BOUND, no release/model/source-SHA fields; (3) HEAD moved d4848f8d -> e0023a98 -> fe4fcce8 and warehouse rows moved (n_main 141,345 -> 141,353). arms.py differs (d654b285 -> e782a10b) by one line in early_season(), which rows_for/score_arm do not call; share.py and playermodel.py are byte-identical to the record's provenance. Shipped MAE therefore moved 3.71988 -> 3.72013, which is data vintage, not code.

Finding and verdict

UPHELD

Review scope

UPHELD: roster3 (six l3 columns over his last three ACT rows) does not beat the shipped projection, and H-224 as filed (age 4+ better by 0.10-0.30) is refuted; reproduced on a fresh snapshot at a later HEAD. WEAKENED on three secondary claims inside the record: 'pooled nil HELD', the age-4+ 'worse than shipped, p=0.065' half that drives next(1), and 'does not improve on any stratum'. Not a PROMOTION.