Is the 13:21Z lead real out of selection, and where does it sit? Four seasons (2019-2022) no H-021 run has scored, each a fresh production refit (season <= S-1), with the same x (8-lag hl2/hl4 geometric pool of the weekly share minus lags.sql season-to-date share).

Owner lane: s-observation

Question

Is the 13:21Z lead real out of selection, and where does it sit? Four seasons (2019-2022) no H-021 run has scored, each a fresh production refit (season <= S-1), with the same x (8-lag hl2/hl4 geometric pool of the weekly share minus lags.sql season-to-date share).

Prediction

Frozen in protocol.json (sha 85d62f8b) at 15:17 MT before any 2019-2022 row was scored: (a) primary target_share_season@hl2->targets pooled 2019-22 in [+0.02,+0.07], outside Bonf-7, positive in >=3 of 4 seasons; (b) >=3 of 4 snap/carry cells outside Bonf-7; (c) target share->fantasy inside its band; (d) decile shape top>bottom, no decile correction worth > 0.005 targets/row, walk-forward nil; (e) HEAD refit of 2023-25 reproduces stored rows. REFUTED if primary <= 0 or inside band.

Finding and verdict

CONFIRMED OUT OF SELECTION and LOCATED; still small. Primary +0.041 on 2019-22 (4/4 seasons), all 7 cells outside Bonf-7 in both populations. Concentrated in weeks 2-4 (rho ~0.09 pooled, linear bound 0.0053 targets/row = 0.38% of MAE 1.41) vs week 5+ (0.0007). A 10-bin walk-forward correction is nil (+0.0074 vs shuffled +0.0066). MEASURED NOTHING end to end remains the H-021 verdict at +-0.01 pts; the lead is a real early-season observation defect, bounded.

Reasoning

The 13:21Z lead was a post-hoc selection from 36 cells on the rows it was found on; this is its first out-of-selection test, with a frozen prediction, plus the weak joints it named (nonlinear use, HEAD refit).

Next action

Weeks 2-4 only, frozen: x_prior = last-season share (final 4 games) - season_td vs the hl2 pool, to separate prior-season carry-in from recency weighting; and whether any shipped feature (kalman_share) already carries last season at week 2 and what the tree does with it. Worth <= 0.005 targets/row linearly, so the value is the defect, not the points.