The 21:17Z lead (hl2 pool minus season-to-date share correlates with the shipped targets residual, concentrated in weeks 2-4) mixes two things in one x: last season's share entering the pool, and recency weighting inside the season. Which one does the data carry, and is the information already in the served row?

Owner lane: construction

Question

The 21:17Z lead (hl2 pool minus season-to-date share correlates with the shipped targets residual, concentrated in weeks 2-4) mixes two things in one x: last season's share entering the pool, and recency weighting inside the season. Which one does the data carry, and is the information already in the served row?

Prediction

Frozen in protocol.json (sha 1ac5ca37) at 17:17 MT before any decomposed x was computed: P1 C1 (x_prior_f4 -> targets) positive and outside Bonf-6 in both populations, and partial corr of x_hl2 given x_prior_f4 <= 0.5x raw; P2 C4 (x_within, wk 3-4) inside band both populations; P3 C1 survives partialling on served target_share_l3_est + target_share_season (CI excludes 0 pooled); P4 C5 positive, C6 positive and smaller than C1; P5 (descriptive) C1 at week 4 >= week 2. REFUTED if C1 <= 0 or inside band in both populations while x_hl2 stays outside.

Finding and verdict

REFINED: the 21:17Z week 2-4 lead is under-shrinkage of the season-to-date target share, NOT prior-season carry-in and NOT recency weighting. Frozen primary held (C1 outside Bonf-6 in both populations), but a post-hoc amendment registered before it ran shows C1 is a proxy: last season adds +0.028 given f (inside band), while -f alone carries +0.111 pooled (both populations outside band), with a linear bound of 0.0093 targets/row at weeks 2-4, twice the 21:17Z object's 0.0047 on the same rows. The served row carries what the fit needs (F1 nil). H-021 end to end stays MEASURED NOTHING at +-0.01 pts; this is a located fit defect, bounded, not a window finding.

Reasoning

The 21:17Z record named this exact rival as its next step. A recency-window change (H-021's premise) and a shrinkage change make different fixes; filing the wrong mechanism would send the next arm to the wrong object.

Next action

Frozen, prospective: (1) 2026 weeks 2-4 ISSUED lines (aar_forecasts vs aar_outcomes, the shipped object) -- corr(-f | position, targets residual) > 0; not before 2026-10-06, when week 4 settles. (2) Week 5+ on the stored rows: does -f | position decay to nil, or persist (the 21:17Z +0.025/+0.043 residue)? (3) Position-mean vs room-complement shrinkage target. (4) Week 1 and rookies, which this population excludes.