With the replica finally exact (order-preserving TEMP shadow, the H-274/H8 workaround), does D-209's EWMA(hl=2) beating flat l3 survive end to end for targets_l3 and carries_l3 through the production path, and if not, is the recency signal lost in the chain or already carried by the line?

Owner lane: forecast

Question

With the replica finally exact (order-preserving TEMP shadow, the H-274/H8 workaround), does D-209's EWMA(hl=2) beating flat l3 survive end to end for targets_l3 and carries_l3 through the production path, and if not, is the recency signal lost in the chain or already carried by the line?

Prediction

Registered 07:29Z in protocol.json before any fit: (a) order-preserving flat replica bit-identical on every record in all three seasons; (b) P pooled fantasy-point MAE difference between -0.008 and +0.004, all Bonferroni-6 intervals contain zero on all three endpoints (nil); (c) single shuffled-label delta within +-0.002 on fantasy points. Follow-up registered 09:1xZ (followup-protocol.json) before its run: shuffled draw inside central 95% of 2,000 permutations; model-free EWMA-minus-flat negative on targets and carries with intervals excluding zero, under 0.10. Addendum (followup2-protocol.json), registered after follow-up 1 was read: REDUNDANT, |corr(x, shipped residual)| < 0.03 and walk-forward MAE delta within +-0.01.

Finding and verdict

MEASURED NOTHING end to end, and the reason is now known. Exact replica, production path, 20,523 player-weeks 2023-25: EWMA(hl=2) on targets_l3 + carries_l3 moves fantasy MAE +0.0075 [Bonf-6 -0.0065, +0.0215], targets +0.0025, carries -0.0003, every interval spanning zero. The D-209 recency signal exists before the model (-0.052 targets, -0.033 carries model-free) and is REDUNDANT with what the line already reads (corr with the shipped residual 0.003 / 0.026). A composition change to a raw-count window that the line reads through a filtered share cannot transmit. Resolution note: this refit moves forecasts 0.49 pts/row with no information change, so composition arms on this layer resolve about +-0.01 pts MAE, not less.

Reasoning

Closes 2 of H-021's 35 windows with an exact replica for the first time (09-13, 09-24 and 09-25T0530Z runs were voided by the replica). Gives the other 33 a cheap screen: the model-free gain and corr(x, shipped residual) cost seconds and need no refit, and a window whose x does not correlate with the shipped residual cannot pay end to end.

Next action

H-021 stays OPEN for 33 windows. Screen each with model-free gain + corr(x, shipped residual) before any refit; run the end-to-end arm only where the correlation is material. T (plays_l3, pass_rate_l3) needs a MATCHED_BOTH_BALL design because the opponent's copy changes too.