Review of 2026-09-25T2117Z-h021-share-recency-heldout

Owner lane: Not recorded

What held

The primary lead is real out of selection: 0.0413 on four seasons no H-021 run had scored, positive in 4/4 seasons, player-clustered t 4.23 (CI [0.022,0.061]) against a calibrated null, and the same on replay (t 4.53). Target-share and snap-share recency on the targets endpoint (4 of 7 cells) survive Bonf-7 under valid clustering in both populations. The record froze its prediction, disclosed 2022 as weak, disclosed the post-hoc breakdown, the failed first analyze attempt, the 2025 warehouse drift and the unstable position split, and its size claim (0.0011 targets/row overall, 0.0053 at weeks 2-4) is honest. Stored-row replay and baseline identity check out to 0.0.

What did not hold

(1) "all 7 cells outside Bonf-7 in both populations": 4 of 7 under player-clustered inference; the carry cells and the fantasy cell do not clear, and (c) "the 13:21Z does not reach fantasy was a Bonf-36 band artefact" is not supported. (2) Prediction (b) scored RIGHT only under the row-permutation band. (3) "located to weeks 2-4" rests on one uncorrected post-hoc split, RB persists to week 5+, and the hl2 cells' covariance is mostly between player-season, not early-season timing. (4) "walk-forward nil" is a power statement (expected gain about 0.0013 vs noise 0.0066), not a measured absence. (5) The record's null (row permutation within season) is the wrong exchangeability unit; the same mistake the prior referee found on the 13:21Z screen, repeated in the record built to confirm it.

Finding and verdict

WEAKENED

Review scope

WEAKENED, not refuted. UPHELD: the primary target-share recency residual on the targets endpoint is real out of selection (2019-22 and 2023-25, player-clustered t 4.2-4.5), and small (bound 0.0011 targets/row; 0.0053 at weeks 2-4); target-share and snap-share hl2/hl4 -> targets survive Bonf-7. NOT UPHELD: 7/7 cells (4 of 7), carry-share recency (t 1.9/2.7 held, 1.5/2.5 replay, prior referee also refused it), target share -> fantasy points, the weeks 2-4 location and its carry-in mechanism (RB persists; hl2 covariance mostly between player-season), and walk-forward nil as evidence. The 2022 season alone is nil (t 0.5). FINDINGS row 165's exception is therefore not withdrawn, but it should name 4 cells and cite the clustered interval, not 7 and the row-permutation band.

Next action

(1) Any H-021 or share-window claim uses player-clustered (or block-permutation) inference and states the null's size control; row-permutation Bonf bands are retired for repeated-player rows. (2) The between-player-season part of the hl2 signal is a role-trend feature question (recent share minus season share as a level, all season), separable from prior-season carry-in: test x within weeks 2-4 vs a season-long x on 2019-22 with player-block permutation. (3) Bind snapshot bytes (sha256 of the .duckdb) in every screen record; the file this record names was deleted. (4) A loss-matched walk-forward (LAD scored in absolute error, or OLS in squared error) with a planted control at the observed effect size (rho 0.04), not 0.19.