Review of 2026-09-24T0222Z-h184-band-centre

Owner lane: Not recorded

What held

Premise check (board centre = line x factor_for; SHIPPED_CURVE fitted on the unblended walk-forward centre; ledger-is-unblended clause false at HEAD) is confirmed in code. The primary and every secondary/control number reproduce as values in both seasons. The retired-constant-band finding on live weeks 1-2 reproduces and is the most useful line in the record: 100 percent of pre-kickoff banded rows with a positive prediction on the 09-10 and 09-11 filings carry 0.45x/1.75x.

Finding and verdict

WEAKENED

Review scope

SPLIT. SURVIVES: the direction of the headline. Blending the centre does not hurt the SHIPPED_CURVE band and refitting the curve on the blend does not help it, in the one genuinely held-out season (2025): secondary B-refit - B-ship pinball +0.0013 [-0.0053, +0.0076], cover -0.0028 [-0.0061, +0.0004]. The shuffle falsifier fires (+2.29 [2.13, 2.44]) and I added a known-size centre-mismatch control that also fires. DOES NOT SURVIVE AS WORDED: (a) the frozen PRIMARY contrast (B-ship - U-ship pinball -0.042) does not measure 'the cost of the centre mismatch': B is a better-ordered centre (D-80), so its pinball is lower for reasons that are the centre's accuracy, not the band's fit to it; the record's own mechanism sentence says so, yet the verdict and prediction_scored read the primary as 'the band's favour'. The mismatch cost is bounded only by the secondary, and that is nil-not-proven (CI half-width 0.0064 pts of 1.52 = 0.4%) on ONE season; the 2024 secondary is +0.0206 [+0.0078, +0.0326], which the record attributes to SHIPPED_CURVE having seen 2024 but which is not separable from a real refit cost on the data available. (b) 'costs the SERVED band nothing' pools four positions, but since 2026-09-23 (forecast.py:542) only QB is served from shipped_band whenever a sim band exists; RB/WR/TE are served from sim_p10/sim_p90 and the record says it did not replay them. The QB-only slice holds (B-ship - U-ship -0.110 [-0.154,-0.064]; secondary +0.029 [-0.002,+0.059]) but the record never reports it and the refit QB band is 0.75 pts NARROWER at equal pinball. (c) the live-2026 extension is ad hoc SQL outside the frozen runner (added after the historical run, disclosed) and, re-derived, is correct as far as it goes -- but the 09-15 fitted-curve filing has since settled for week 2 and covers 0.7535 (n 353), below every one of the 18 walk-forward 2025 weeks (0.837-0.902), so 'the curve over-covers, ceiling slightly high, is the next question' does not carry to live.

Next action

(1) File the QB-only served-band contrast and refit-narrowing as the served-position statement of H-184; retire the pooled four-position wording. (2) Grade the 09-15 SHIPPED_CURVE week-2 under-cover (0.7535) as its own question against the historical 0.837-0.902 weekly range once weeks 3-4 settle; separate the FantasyPros-covered rows (vintage: live vs archival) from the rest. (3) Make fit_shipped's position order deterministic so the result sha is a stable identity (a mechanical crumb for the mice, not filed by this lane). (4) A mismatch-cost measurement should difference two bands on the SAME centre (B-ship vs B-refit) and treat B-ship - U-ship as centre quality.