Review of 2026-09-26T0734Z-h193-update-calibration

Owner lane: Not recorded

Finding and verdict

UPHELD

Review scope

SURVIVES: exact, field-for-field reproduction of the analyze step from the frozen preds_wf.csv.gz and a freshly-taken warehouse snapshot; production entry point used correctly (wf_std.py -> PM.fit strict season<S -> share.feature_sql -> PM.lines_for); no leakage in the fit window or the anchor-recentring path; MAE and MSE kept separate rather than conflated; the 6-member confirmatory family and its post-hoc additions are disclosed, not folded in silently; both named shuffle controls (E5a, E5b) are real and collapse as required; the declared falsifier does not trigger, and the record's own pre-registered verdict rule (not a post-hoc restated bar) produces exactly the filed PARTIAL verdict when checked mechanically against the reproduced numbers. DOES NOT FULLY SURVIVE, but does not change the verdict: (1) FantasyPros' post-kickoff archival fetch clock (aar.py part-oracle caveat, disclosed in the record's own weak_joint_for_referee) means its apparent underreaction could be partly late-week information the model could not have had -- the record already flags this rather than hiding it; (2) the fpapi projections table read from a live/fresh snapshot is not itself hash-bound as an input artifact, so a change to that table between now and a future review would not be caught by hash comparison alone -- today it reproduced exactly, but this is a real gap in the evidence contract's artifact list, not a substantive result issue.

Next action

The record's own next-steps are sound: (1) sim.py/prop-probability/kernel consumers that read the line as a MEAN inherit the 15% over-sized WR/TE revision the board's 0.40 blend hides -- a mean-consumer walk-forward (prop Brier, sim totals) belongs to those layers, not this one; (2) H-190's consensus-weight-0.95 finding and this record's blend calibration (E9, post-hoc) point the same direction and a weight-grid follow-up on b_R would tie them together; (3) the opponent decomposition of the revision (both_sides) remains open and would need a separate MATCHED_BOTH_BALL design with an explicit defense-side term.