Review of 2026-09-24T0229Z-h185-prop-band-early-weeks

Owner lane: Not recorded

What held

Premise checks are honest: TRAIN_MIN_WEEK=4 and the first-stake-is-week-5 point are real; the frozen prereg is hashed by the runner; all 15 endpoints are reported; prediction_scored says MIXED and states the formal refutation bar was not met. The rush_yards early centre (+1.04 yd, 5/5 seasons) reproduces and is the one early-week signal with a season-level t of 3.1 (still exploratory, not family-level). The band-widening premise (early residuals wider) does not appear at any level tested. The week-banded arm is inert relative to its shuffled null and falls back to A for QBs by construction.

What did not hold

Headline verdict 'REFUTED AS FILED' is stronger than its own prediction_scored ('formal refutation bar not met'); the QB per-fold sign sentence is wrong against its own data; the 'surviving' all-week over-pricing rests on a block bootstrap that ignores that calibration is per season.

Finding and verdict

WEAKENED

Review scope

SURVIVES: the reproduced values (every family endpoint within drift; no sign or exclusion flips), the null on weeks-1-3 width (E2 spans 0 on all five), rush_yards centre E>L in 5/5 seasons, and the harness controls (K1 in-sample <=0.0072, K3 unchanged). DOES NOT SURVIVE: (a) 'H-185 REFUTED AS FILED' for the QB throwing-yards centre and width -- the interval is wide (E-L -9.6 [-23.9,+3.6] at 99.67%, [-19.5,+0.14] at 95%; 459 early rows) and includes the filed +2..+12 yd claim, so the run is UNRESOLVED for QBs, not a refutation; (b) the record's own per-fold sign statement for QB throwing yards ('early LOWER in 4/5 folds: 2021,2022,2023,2025') is contradicted by its own by_fold block -- early-minus-late centre is -34.5, -20.4, -9.4, +8.4, +6.8 for 2021..2025: LOWER in 3/5, HIGHER in the two most recent seasons, and 2024/2025 sit inside the +2..+12 yd the claim predicted; (c) 'the finding that survives: OVER over-pricing in every bucket, all five quantities clear of 0 at 99.67%' -- the interval comes from an 85-block season-week bootstrap that treats blocks as exchangeable although every row of a season shares ONE calibration (five independent calibrations in total); on the five per-season late-bucket biases (t, 4 df) none clears 0 at 95% (carries p 0.075, rush_yards 0.056, QB throwing yards 0.146, receptions 0.107, rec_yards 0.109) and seasons 2022/2024 are at or below 0 on carries, rush_yards, receptions, rec_yards.

Next action

(1) Restate the QB throwing-yards early-week question as UNRESOLVED (459 early rows, 5 seasons, sign flips in 2024/2025) and keep it OPEN; the honest prospective test is 2026 weeks 1-3 QBs with real lines. (2) Any statement about over-pricing at week 5 must be made at the season/calibration-vintage unit, with the recentred-centre control alongside; route it with H-041/H-180, not as an H-185 result. (3) The record's per_fold_E1_sign sentence should be corrected by an addendum, not a rewrite.