Review of 2026-09-25T1321Z-h021-last6-screen

Owner lane: Not recorded

What held

E1 reproduces to 1e-12 and its direction is real (10 of 12 cells better than flat, CIs excluding zero; td_oe hl2 and hurry hl2 worse). The three non-share windows (td_oe_l8, hurry_delta_l8, own_scramble_rate_l4) are nil on E2 under both row-permutation and player-clustered inference (max |t| 1.35), for both hurry definitions. Target-share and snap-share recency on targets is a real residual: three cells clear Bonf-36 under player clustering and replicate in sign in 3/3 seasons, and the record flagged it, the linear bound, and E3's blindness itself.

What did not hold

(1) 'MEASURED NOTHING end to end' for all 6 is contradicted by its own lead and by A2c: a matched-loss add on targets gains ~0.3% squared error, out of fold, control functioning. (2) 'E3_valid_candidates_significantly_better: 0' and 'F4 ... worse on all 6' are artefacts of scoring an OLS slope by absolute error on peaked, heavy-tailed residuals; they carry no information and should not have been counted toward the 24-of-35 nil tally. (3) The carry-share lead is not supported under player clustering. (4) The inference for E2 ignored player clustering, inflating the breach count from 3 to 6. (5) The hurry result is about a feature definition that the warehouse no longer holds and that an uncommitted edit is replacing; the record does not name the dependency. None of this makes the size large: the recency-contrast lead is ~0.3% of squared error on targets and nothing on absolute error or on fantasy points (E2 fantasy cluster t <= 2.12).

Finding and verdict

WEAKENED

Review scope

WEAKENED, not refuted. Upheld: nil for td_oe_l8, hurry_delta_l8 (committed definition, and current definition as sensitivity) and own_scramble_rate_l4 on the E2 correlation, and E1's direction. Not upheld: (a) 'MEASURED NOTHING' for the three season shares - target_share_season and snap_pct_season (hl2) on the targets endpoint carry a small out-of-fold squared-error gain the shipped line does not contain; (b) the carry_share_season->carries lead; (c) the E3/F4 nulls as evidence of anything; (d) the '24 of 35 nil' tally to the extent it counts E3. The gain is small (about 0.3% of targets squared error, no measurable absolute-error gain, no fantasy-point gain) and was found on rows that also selected it, so it is a lead to confirm on 2026 weeks with a matched-loss instrument, not a finding to ship. Not a claim about the 11 team windows or about the 2026 live rows. An unexplained residual of this size is information not yet found, not a floor.

Next action

(1) Any further H-021 screen uses a loss-matched instrument (OLS fit scored in squared error, or LAD scored in absolute error) and player-clustered E2; drop the MAE-scored OLS add. (2) Confirm the target-share/snap-share recency contrast on the targets endpoint on 2026 weeks 3+ (aar_forecasts vs aar_outcomes), player-clustered; carry share is not a candidate. (3) Bind the snapshot: record sha256 of the .duckdb bytes and the tempo_week/tempo_delta table hashes for any screen that reads them; the file the record names has been overwritten. (4) Commit or revert the playerweek/tempo.py edit and say which hurry_delta_l8 production serves; the stored shipped rows and every H-021 hurry number use the committed one.