Review of 2026-09-23T0903Z-h202-concentration-week-estimand

Owner lane: Not recorded

Reproduction

REFUSED. Bound bytes verified: record 9b88e05b..., protocol = runner 12faa135... (equal to prereg copy and to run/), result out.json a43f4744..., share.py 37758ac8... (matches HEAD). concentration.py CHANGED: bound 79c79513..., disk 7d541eb0... (C-13 d7e40a23 added a timezone('UTC', now()) AS ingested_at column to the team_concentration SQL; values unchanged, replica max abs diff 2.2e-16 on 229/227/226 keys in every run). arms.py also changed since fb922957 (a default-argument constant in early_season, not on this path). Not bound: the runner freezes `git archive HEAD` (now moving) whereas the record names fb922957, which the reproduction pins; the source snapshot sci-h202.duckdb hash appears in no field and the warehouse has moved (shipped MAE 3.71988 -> 3.72122 here, 3.71932 in the earlier referee run); baseline.status NOT_BOUND (refit per fold, PW_FITCACHE=0, no release/model/source-SHA); information_set NOT_VERIFIED. The run-2 stdout was lost, so the record's numbers rest on out.json alone.

Finding and verdict

WEAKENED

Review scope

UPHELD: Part A enumeration reproduces to four decimals (WR squad gap +0.0483, RB carry +0.0821, group WR +0.0843, RB +0.1051; season leader's own weekly mean 0.2505 WR / 0.5376 RB, so 41% / 58% of the gap is his absences, A3 FAILED as recorded); shipped_again 0.0 max abs; manual vs score_arm 5.0e-5; replica of CC.targets 2.2e-16; fp_l3 naive loses to shipped by +0.568 [0.498,0.633]. The headline point estimate reproduces in sign and size: squad_week - squad fantasy MAE -0.01188 (record -0.01121) on a fresh snapshot with the tree pinned to fb922957; the same contrast on an earlier unfiled referee run of the same runner: -0.00995. The squad-CARRY split gain of the week-keyed arms reproduces in all three vintages (squad_week-squad -0.00239, group_week-shipped -0.00280, Bonf6 excluding zero each time). WEAKENED: (1) The claim that the headline clears Bonferroni-6 is a margin of about 0.001 pt and does not hold across warehouse vintages of the same code: record hi -0.0002 (shipped MAE 3.71988); my rerun hi -0.00077 (3.72122); the earlier unfiled referee rerun of the identical pinned runner (3.71932) hi +0.00119, spanning zero. The 95% interval excludes zero in all three ([-0.0184,-0.0013] to [-0.0205,-0.0033]) but the pre-registered Bonf-6 bar is crossed in two of three. It survives the cluster-unit sensitivity on the vintage where it holds (week -0.00026, team-season -0.00037), yet the vintage moves the bound more than the cluster does. The squad-T split is the same: the record's Bonf6 upper bound is +0.00008, mine +0.00017, the other run -0.00038. (2) The comparator is not production. squad_week - squad = -0.0119 is the difference of two arms neither of which is distinguishable from shipped: squad_week - shipped +0.0029 [-0.0144,+0.0197], squad - shipped +0.0148 [-0.0032,+0.0334]. Nothing beats shipped on fantasy MAE, which the record says. 'Recovers ~70% of H-109's MAE loss' is a ratio of two estimates each spanning zero (0.0148 -> 0.0029 is 80% here). (3) 'It is a sharpening LEVEL, not team information' is an inference from a control that cannot separate the two: squad_week_shuf - squad_week is +0.0014 [-0.0078,+0.0112] Bonf6 (other run +0.0018 [-0.0089,+0.0121]), a half-width about equal to the 0.0119 effect, and squad_week_shuf - squad is -0.0105 [-0.0236,+0.0041]. The shuffle does not show the gain is level; it fails to show it is not. A level-matched control (squad scaled to squad_week's mean exponent) was not run. (4) The estimand story is not isolated: squad -> squad_week changes both the time axis and the level of the solved exponent (median 0.93 -> 1.24), so the arm cannot distinguish 'the right unit' from 'more sharpening'. The record's own read says the latter; the question line 'does re-keying to the per-week estimand move fantasy MAE' is answered only as 'not against shipped'. Effect sits in RB (-0.0245) and 2025/2023/2024 point estimates are all negative but each spans zero on its own. Do-not-ship stands on no demonstrated gain against shipped; the carry-only re-key remains a hypothesis for held-out confirmation. Not a PROMOTION.