Review of 2026-09-25T1221Z-h021-pbp-window-screen

Owner lane: Not recorded

What held

Reproduces numerically on a rebuilt snapshot. E1 gain over flat4 is real and 16/16 under either loss and either clustering unit. No window has E2 outside Bonf-48 (pooled) or outside a season-stratified Bonf-144 band except one edge cell; no window has a winsorised or flat8 linear add that beats shipped on the windows the planted control can see; a nonlinear check finds nothing that survives correction. The nil is a nil against a linear-plus-shallow-tree add of these 16 EWMA increments on these 20,523 rows, and the record scoped it that way (weak_joint_for_referee 1-4, all four confirmed as real limits).

What did not hold

The `mechanism` paragraph ('the recency-weighted pool is a better estimate of the rate itself') is contradicted by S2: the gain is pool depth, and hl-2 weighting is worse than an equal-weight 8-row pool on 13 of 16 windows. The `why` line ('E3 valid on 9') understated the repair available: 13 windows after winsorising. Neither changes the verdict on 'does not reach the line'.

Finding and verdict

UPHELD

Review scope

UPHELD for 'the 16 EWMA-hl2 recompositions of player_pbp_l4 add nothing measurable to the shipped fantasy line' (E2 linear, winsorised linear E3, and a shallow nonlinear check), scoped to rows where both pools exist. NOT UPHELD as mechanism language: the E1 gain over flat4 is pool depth, not recency. Not a claim about the 17 unscreened H-021 windows, the coverage-extension rows, or any other half-life/depth. 'MEASURED NOTHING' is a nil with stated power (planted 0.05*sd leak seen on 13/16 linear windows; pooled E2 floor about 0.02 in corr, R^2 ~ 0.0004 of residual variance) and is not evidence that the gap cannot close.

Next action

(1) If H-021 continues on the pbp windows, the candidate E1 ranks first is a flat 8-row pool, not hl-2; screen it (E2 above says nothing there) before more EWMA windows. (2) Read the coverage-extension rows (up to 974) as their own contrast: x is 0 there by design. (3) Bind the snapshot bytes: the file the record names has been overwritten twice since.