Review of 2026-09-25T0326Z-h089-served-decision

Owner lane: Not recorded

What held

The direction and the size on neutral scales survive every attack that could kill it: stale served row orders close same-position pairs worse than the trained row by about 1.2-1.3pp accuracy and +0.15 pts regret per pair (product), on the record's selection, on production latest_rows' involved-row selection, and on a symmetric pair set, with permutation p 0.003-0.010. The post-fit arithmetic reproduces byte for byte. The disclosure of weak joints (a), (b), (f) is accurate. H-089's 2-5pp further-cost prediction failing on magnitude is correct.

What did not hold

(1) The frozen PRIMARY 'excludes zero' is not stable under a refit (upper bound +0.04/+0.06 on fresh runs) and the raw primary spans zero; the honest primary statement is a point estimate of about -0.8pp with an interval touching 0. (2) 'SEL_T under-states the served cost by half': the neutral-scale figure is 1.2-1.3pp against 0.82, a third, and SEL_S (-1.5) is inflated by conditioning on the served arm's own closeness, so 'the selection the consumer faces' number is the least neutral of the three. (3) The S arm is a reconstruction (previous played row) and not the production selection; it happens not to matter for this endpoint (unlike H-222) but the record did not test that. (4) The 2025 season (the only one after the last fit window edge and the most recent) is -0.5pp on SEL_C with an interval spanning 0, and the effect halves relative to 2023-24; a dated cost claim about the current board rests on one season. (5) Consensus margins inherit an unverified fpapi vintage.

Finding and verdict

WEAKENED

Review scope

SURVIVES: direction and the ~1.2-1.3pp product accuracy cost / +0.15 pts regret on pair sets not conditioned on either arm alone, under both the record's previous-played-row arm and the production involved-row selection; post-fit arithmetic reproduces byte-identically. DOES NOT SURVIVE: the frozen primary's exclusion of zero as a stable finding (spans 0 on a fresh refit of the same runner at two bootstrap settings); the 'SEL_T understates by half / SEL_S is the cost the manager faces' framing; any statement about 2025-and-later that ignores the 2025 attenuation. UNTESTED: grading of issued aar_forecasts rows (no 2026 arm), a random-other-week prior-row intervention separating 'stale' from 'different', fpapi pre-kickoff vintage.

Next action

(1) File an addendum, do not rewrite: primary point -0.8pp, interval touching 0; neutral-scale (SEL_M/SEL_C) -1.2 to -1.3pp. (2) A decision gate on the repair (H-222's proposal) should use SEL_C/SEL_M-style pairs, not SEL_S, and require the interval on 2024-25 pooled, since 2025 alone is -0.5pp. (3) Bind a byte-identical snapshot and fitted-row cache in input_artifacts before any promotion; the refit is not byte-reproducible (H8). (4) Prospective grade of issued 2026 aar_forecasts vs fpapi_live once 4+ weeks settle.