Review of 2026-09-23T0611Z-h109-concentration-denominator

Owner lane: Not recorded

Reproduction

REFUSED. Bound bytes verified: record dab83d4b..., protocol 6e0be4ba..., runner 3da30295..., result 6750c1f4..., share.py 37758ac8... (matches HEAD). concentration.py CHANGED: bound 79c79513..., disk/HEAD 7d541eb0... (C-13 d7e40a23 added a timezone('UTC', now()) AS ingested_at column to the team_concentration SQL; values unchanged, verified by replica max abs diff 2.2e-16 on every key in the reproduction). Not bound: the runner freezes `git archive HEAD`, which is now 276ecbcd, and HEAD is 77 playerweek/ files past aa0f6737; the source snapshot sci-h109.duckdb is gone and its hash is in no field; baseline.status NOT_BOUND (refit per fold, PW_FITCACHE=0, no release/model/source-SHA); information_set NOT_VERIFIED. The reproduction pins the tree to aa0f6737 but runs on a new warehouse vintage: shipped MAE 3.72013 vs 3.71988 (+0.00025), so warehouse rows moved slightly even for 2023-2025. That vintage drift alone moved the MAE lower bound from +0.00018 to -0.00036.

Finding and verdict

WEAKENED

Review scope

UPHELD: the premise (squad-denominated WR target 0.228-0.234 vs shipped 0.384-0.394; target-solve fallback 0.254->0.040; median solved target exponent 1.93->0.93) and the direction of the consequence (squad MAE +0.0155 pts, squad-T split +0.0018, oracle_serve not better than shipped) reproduce on a fresh snapshot with the tree pinned to aa0f6737: squad MAE +0.01547 vs record +0.01557, squad-T +0.00182 vs +0.00188, squad_shuffled +0.01948 vs +0.01942, oracle +0.0092 vs +0.0094; shipped_again 0.0, manual vs score_arm 5.0e-5. WEAKENED: (1) the record's REFUTATION CLAUSE (squad worse on fantasy MAE with Bonferroni-3 excluding zero) rests on a lower bound of +0.00018 and does NOT survive on the same code and rows: reproduction Bonf3 [-0.00036,+0.03186]; week-clustered [-0.00030,+0.03281]; team-season-clustered [-0.00217,+0.03397]. The squad-T interval spans zero at both HEAD runs in the record itself, so 'worse on the squad-T split by 0.002-0.004' is a claim about afb1d233 only; at the headline tree it is +0.00188 [-0.00023,+0.0039]. The evidence is 'no gain shown, harm not established', not 'harm established'. Prereg P3 (nil, |d|<0.005) is not confirmed either: point estimate 0.0155 exceeds 0.005 and the interval spans both 0 and 0.005. (2) The MAE cost is not a team-identity effect: squad_shuffled costs the same (+0.0195; squad minus shuffled -0.0040 [-0.0126,+0.0043]). It is carried by 2024 (+0.030) and RB (+0.034); 2023 is -0.0006. The record says 'worse in both code states' but the two runs are the same rows and the same season fold, so it is one sample twice, not a replication. (3) The mechanism sentence 'the target is operating as a fitted sharpening knob, not a measurement of concentration' is not identified by the oracle arm: oracle_serve fits calibration on the shrunk trailing squad target (KEEP=0.35) and serves the unshrunk realised SEASON leader share, so fit and serve targets differ by design, and a season leader's share is not the expected share of the week's model-picked top WR that alpha_for_subset solves for (the H-202 estimand question the record defers). Oracle worse-than-shipped therefore does not separate 'wrong-unit target compensates downstream' from 'season-leader share is the wrong estimand at either denominator'. (4) The decision endpoint (start/sit) is not measured, as the record says. Do-not-ship of the relabel alone stands on no demonstrated gain. Not a PROMOTION.