Review of 2026-09-23T1726Z-h203-band-spread-by-defence

Owner lane: Not recorded

Finding and verdict

WEAKENED

Review scope

The record's numbers reproduce byte-for-byte. What does not survive is the headline reading of the RB row: (1) its interval excludes zero only under a game-cluster bootstrap, while the treatment (a defence's trailing-points tercile) is nearly constant within a defence-season, so the independent unit is ~96 defence-seasons, not 816 games; at that unit the RB 98.75% interval is [-0.07,+5.28] pp and a defence-level permutation p is 0.016 against the family's Bonferroni-4 threshold 0.0125. (2) The record's own post-hoc itt split halves it (+1.33 [-1.15,+3.90]). (3) The object measured is the FALLBACK curve for RB/WR/TE, not what forecast.py serves for them. QB, WR, TE nulls stand and are not weakened: none of them exceeds its permutation null.

Next action

The only claim worth a further frozen test is the RB one, and on the SERVED band: sim_world sim_p10/sim_p90 for RB, soft-minus-hard exceedance at fixed level, cluster = defence-season, family fixed before the run, an itt-stratified arm in the family. H-255 (itt-keyed shipped curve) already targets the fallback curve, so it inherits the same premise gap.