Review of 2026-09-23T0533Z-h212-factors-band-input

Owner lane: Not recorded

Reproduction

REFUSED. Source bytes verified unchanged on disk: record 2ee1f411..., protocol bf3d1f0c..., runner 4943f392..., result 042129ee..., allocation.py 2b638d83..., lags.sql 499592a9..., playermodel.py eeeb6a3c... (provenance). Not bound: (1) the input warehouse: the source snapshot h212-band.duckdb is gone and its hash is in no field; (2) baseline.status NOT_BOUND, no release/model/source-SHA fields, no saved model, each fold refits at PW_FITCACHE=0; (3) information_set NOT_VERIFIED, no cutoff or receipt clocks; (4) HEAD moved 140f427a -> c25a2130 and warehouse rows moved: shipped BASE MAE 3.72132 -> 3.72293 and rows with no board reconstruction 513/512 -> 528/533. share.py and arms.py changed only in the S-15 scoring and a calendar guard that rows_for/lines_for do not call (arms.py early_season one line); allocation.py, lags.sql and playermodel.py are byte-identical. The drift is data vintage, not code.

Finding and verdict

UPHELD

Review scope

UPHELD: the served FACTORS band input is not the fitted unit (walk-forward 5.0-8.6% of bands overall, R1 11.5-24.0%, reproduced to within 1-4 rows of 298/482 and 285/465), and feeding it moves pooled fantasy MAE by +0.0006 [-0.0040,+0.0053] (record +0.0008 [-0.0038,+0.0054]), same sign, same interval; the replica and B controls reproduce exactly. WEAKENED on five secondary claims inside the record: (1) 'MEASURED NOTHING' is a power statement, not an absence: 89% of rows are exactly unchanged, and the 1,201-1,206 changed rows carry +0.015/+0.017 [-0.030,+0.057]/[-0.027,+0.060], a half-width of 0.043 pts/row that cannot exclude a harm as large as the shuffle's; (2) FACTORS were fitted on 2019-2024 (allocation.py:41-45, refit_allocation.py HOLDOUT=2025), so 2024 is in-sample for the constants and only 2025 is held out; the 2025-only contrast is +0.0026 [-0.0008,+0.0061] (repro +0.0021 [-0.0015,+0.0058]), positive both times, the 2024 half is -0.0010 and is the in-sample one; the record does not say so; (3) the shuffle control that shows the harness sees the band clears zero by +0.0007 in the record and by +0.00006 in the reproduction: the pooled resolution is about the size of the effect being called nil; (4) prediction_scored says 'HELD' though R2 carries fell below the stated 5% floor (3.5-4.1%) and R1 targets exceeded 15%; (5) the design isolates FACTORS only, LEAD_FACTORS, the other consumers of the rewritten l3 keys and any median map were left at fitted-unit values, so 'the mismatch costs nothing' is a claim about one consumer. Part A (live R1) is a dated observation and is not reproducible: the board has since flipped to 2026 (450 served rows against 633). Not a PROMOTION.