Review of 2026-09-23T1219Z-h180-prop-band-conditioning

Owner lane: Not recorded

Reproduction

REFUSED. Verified byte-for-byte: source record 560ae718..., protocol cf5baeac..., runner 3e9768b1..., result out.json 31d1090f.... Not bound or changed: (1) input_artifacts is []; the decision_h180.duckdb snapshot is not on disk, and one of its two attempts warned MAY BE TORN (row census reconciled by the author); no snapshot hash is recorded anywhere. Fresh snapshot referee_h180.duckdb holds week 3 played and the 09-25 ledger, so row counts differ by 3 (4,580 vs 4,577 low-p rows; 22,128 vs 22,122 overall). (2) playerweek/playprob.py bytes changed since binding, 3d838d69 -> de83ca65 (PW-148 adds a `roster` argument to fit(), default 'act' preserves the shipped population; diff read). playermodel.py eeeb6a3c, props.py 8506d187 and conformal.py 15e45458 are unchanged. (3) The record binds 'HEAD 8692359c', a release id with no model artifact hash, and the evidence-contract fields for baseline release/model/source, information cutoff and receipts are absent. Replay evidence is about the claim, not those bytes.

Finding and verdict

WEAKENED

Review scope

The measurement REPRODUCES (non-exact, fresh snapshot) and its falsifiers hold; what weakens is the benchmark identity and the framing. (1) REPRODUCE: the runner text, changed only in snapshot name and OUT path, on a fresh snapshot gives p<0.95 L0 pooled n 4,580 (record 4,577), bias_A +0.0717 [0.0524,0.0880] (record +0.0720 [0.0527,0.0889]), bias_B +0.0485, A-B +0.0232 [0.0226,0.0238], p>=0.95 A-B +0.0193, Brier A-B lo [+0.0017,+0.0037]; K1 max|bias| 0.0069, K2 0.0068 (record 0.0083/0.0048); per-quantity lo bias_A receptions +0.049, rec_yards +0.054, carries +0.119, rush_yards +0.079. Two fresh runs are byte-identical (sha 362f29d2...). No leakage found: PM.fit and playprob.fit run on season <= S-1, and the calibration season's stored values come from a bundle fitted <= C-1. (2) WEAKENED, BENCHMARK: arm A 'SHIPPED' is a mirror of props.conformal_cal, not the shipped call, and it is not the same object. props.conformal_cal(con,'season <= 2024') reads stored arm_prediction rows (one build, 2026-09-13 22:09Z, fitted_on 'season <= 2023'); the runner refits lines_for live. Bands are NOT equal for any of the five quantities (S1: equal false x5; stored 2025 rush_yards mean 9.3126 vs runner 9.3736; 2024 rec_yards 19.143 vs 18.974). Pricing 2025 (scored season, 4 quantities) with the PRODUCTION bands instead of the mirror: pooled p<0.95 bias 0.0771 (mirror A 0.0892, B 0.0668), so prodA-B is +0.010 at p<0.95 and +0.014 at p>=0.95 against mirror A-B +0.022/+0.023. The band-mismatch component the record puts at ~+0.020 roughly halves under the shipped bands (caveat: B still uses runner-refit values; one season; so it is a sensitivity, not a replacement). The total over-pricing is not halved: 2025 carries lo +0.170 and rush_yards lo +0.118 under production bands, both inside H-180's 9-30 range. So the split 'one third band, two thirds conditional level' is not stable to the band source, and the level share is larger than stated. (3) WEAKENED, FRAMING: 'MAGNITUDE AND SHAPE REFUTED' rests on pooling over quantities; the pooled figure is dominated by rec_yards (2,069 of 4,580 low-p rows, +0.054) while carries and rush_yards, the quantities H-180 named, reach or sit at the low end of its range. The record does say this in its caveat, but the verdict line does not. On concentration, H-180's question, the data say YES: bias_A is +0.072 below p 0.95 and +0.007 above (my bucket-label shuffle: 0.0757 real vs 0.024 mean, 0.036 max over 60 shuffles). Only the band-mismatch part is band-wide. What the record establishes: over-pricing at p<0.95 is real, +0.07 pooled with a CI excluding zero and below 0.09 pooled, concentrated in low p, and mostly not the band/division mismatch. UPHELD parts: mechanism direction, K1/K2, Brier-better-under-B at low p, and the record's own withheld claims.