Review of 2026-09-23T1330Z-h181-prop-position-pooling

Owner lane: Not recorded

Reproduction

REFUSED as exact: source record input_artifacts is [] (the as-run snapshot sobs_h181.duckdb is not bound and no longer the object read), and playerweek/playprob.py differs from the recorded baseline (record 3d838d69..., now de83ca65b074...; PW-148 added an optional roster= argument to playprob.fit whose default 'act' leaves the fitted table unchanged; the runner calls fit through PM.fit with defaults). playermodel eeeb6a3c, props 8506d187, conformal 15e45458 match the record. Protocol sha 9e7788df, runner a56760ff, result 9340ff45, source record 36752783 all match the packet.

Finding and verdict

WEAKENED

Review scope

UPHELD: the numbers reproduce exactly, and 'H-181 magnitudes refuted' survives (RB receiving +0.045/+0.043, QB carries -0.006, none near +0.10..+0.14 on the scored population). WEAKENED: (a) the scored population is all p_play>=0.5 men, while my stratification shows the over-pricing concentrated in the p<0.95 men (RB rec_yards +0.081 n=775, RB receptions +0.097 n=77), so the refutation is of the unstratified claim only and the record never tested the stratum where a 10-13 point figure could live; (b) 'position heterogeneity real' rests on RB receptions (perm p 0.015, Bonferroni-3 0.045) alone among frozen cells: RB rec_yards is p 0.15 by the record's own null, and TE (p 0.005 = the floor of 200 draws) was selected as the largest of ~13 L0 cells after seeing them, Bonferroni ~0.065; (c) QB carries per-position arm P is uninformative: 2025 fell back to A (A-P exactly 0.0), so the primary A-P is a 2024-only number and 'per-position does not reduce |bias| for QB' compares A with a mixture.

Next action

If the position-pooling question is still wanted: freeze p_play stratum (<0.95 vs >=0.95) as a registered factor, use a 2-season calibration window so QB carries can be banded per position, use >=1000 permutation draws, bind the snapshot as an input_artifact, and score against the stored arm_prediction ledger, not only the replay.