Review of 2026-09-25T0921Z-h021-p-ordered-shadow

Owner lane: Not recorded

What held

The end-to-end contrast is exact, deterministic and reproduces byte-for-byte on fresh bytes; the P arm changes fantasy MAE by +0.0075 [Bonf-6 -0.0065,+0.0215], targets +0.0025, carries -0.0003, and an information-free placebo of the same marginal produces the same move (P - placebo CI95 spans zero on all three endpoints). The D-209 model-free gain is real on these rows (-0.052 targets, -0.033 carries, every season, intervals excluding zero) and it does not reach the served line. The replica/order/negative-fixture controls, the retained records, the disclosure of the failed shuffled-label bar and the honest weak_joint_for_referee list are what made this checkable at all. The record's own claim 'composition arms resolve about +-0.01 pts' is supported.

What did not hold

(1) 'REDUNDANT ... cannot pay' for carries: pooled corr(x, shipped residual) is 0.026 with a game-cluster CI [-0.006,+0.057] that straddles the record's own 0.03 bar, and it is a 2025 phenomenon (0.083 in 2025, -0.005 in 2023, +0.006 in 2024). The walk-forward beta was fit on seasons whose corr was ~0, so it cannot show whether a 2025-sized signal is capturable; the fitted betas change sign across seasons (0.162 / -0.017 / 0.002). This is 'no capturable payoff shown', not 'carried by the line'. (2) The 0.03 and +-0.01 bars were written after the corr values were visible (addendum registered after follow-up 1 was read); they are descriptive bars. (3) The shuffled-label control mis-specified: no valid information-free control existed in the record; the placebo above supplies it. (4) 'MEASURED NOTHING' is right for the pooled interval but hides a QB fantasy-point term (+0.041, uncorrected CI excludes zero) that the placebo shows is inside refit noise (+0.015 placebo). (5) Model-free vs shipped naive comparison zero-fills 5.2% NaN rows.

Finding and verdict

WEAKENED

Review scope

SURVIVES the falsifier: exact replica, production-path P vs shipped nil at about +-0.01 pts (Bonf interval spans zero on fp, targets, carries), reproducible on fresh bytes; the +0.0075 fantasy shift is indistinguishable from an information-free refit perturbation (P - placebo CI95 spans zero); D-209's model-free gain is real and is not visible in the served line. DOES NOT SURVIVE: 'redundant' for carries as a stated fact (corr borderline and concentrated in 2025; the walk-forward test cannot see a signal that only appears in the test season); the shuffled-label control as written; the framing 'MEASURED NOTHING end to end, and the reason is now known' (nil is measured; the reason is measured for the perturbation half only). UNTESTED: the other 33 H-021 windows; T (needs MATCHED_BOTH_BALL); grading on issued aar_forecasts; whether the 2025 carries corr replicates in 2026.

Next action

(1) Treat the record's screen ('model-free gain + corr(x, shipped residual) needs no refit') as valid only with a season-stratified corr and a game-cluster interval, and add the placebo refit (one script, ~40s per fit, 3 seasons in parallel) as the information-free control for every H-021 window instead of a shuffled-label MAE. (2) Score the 2025 carries corr on 2026 weeks once 4+ weeks settle (aar_forecasts vs aar_outcomes) before any 'redundant' claim on carries. (3) Bind a byte-identical snapshot in input_artifacts if this record is ever promoted; the execution snapshot is deleted (record retains rows and hashes).