Review of 2026-09-24T0335Z-h222-served-band

Owner lane: Not recorded

What held

Reproduces to the last digit; controls a and c are real and were seen to work; control b is honestly recorded FAILED; premise_check correctly discloses F-08 and the partial CP-H49; the direction, the season stability, the one-sidedness and the QB null all hold under the production rule; the refit-cannot-repair reading is consistent (V/refitT gap +2.17pp).

What did not hold

Its verdict states the board's ceiling is exceeded 14.3% vs 8.3%. That number is the exceedance on a row construction that includes the D-399 defect production already removed and, for RB/WR/TE on the imminent week, on a band the board no longer serves from the curve. On the production selection the gap is 1.9pp on the rows the curve could still serve, and nil for QB. The mechanism paragraph (role-change) and the 'lower bound' claims in weak_joint (b) are not tested against the involved/uninvolved split.

Finding and verdict

WEAKENED

Review scope

SURVIVES: values reproduce exactly (0 of every levels/contrasts entry differs; n 20,066, 1,632 clusters, meta identical) and the direction of the effect. Ceiling exceedance is higher on a stale served row than on the target row, in every season, at RB/WR/TE, with no QB effect. DOES NOT SURVIVE: (a) the headline number as 'the board's' figure: the served arm is not the row latest_rows serves. It takes each man's previous PLAYED row; production takes his last INVOLVED row (playermodel.INVOLVED_WEEK=True, D-399), because an active-but-uninvolved week has a missing player_pbp_l4 join that share.feature_sql COALESCEs to 16 hard zeros, which the tree reads as 'ran no routes' and D-399 was written to remove. The runner reintroduces that defect on 4,391 of 20,066 rows (21.9%). Under the production selection rule (involved week, in-season once 3 weeks are played, else last complete season minus final 2 weeks; 19,404 rows) the ceiling gap falls from +5.98pp to +1.88pp [Bonf-2 +1.57,+2.17], pinball90 +0.217 -> +0.155, rows served <0.5 where target >=3 from 509 to 59; (b) the mechanism 'rows whose last game moved their role' is mostly this artefact: S-T is +25.1pp on the 4,391 rows whose previous played row was uninvolved and +0.56pp [0.24,0.86] on the 15,675 involved ones (partition check: .219*25.05+.781*0.56=5.9); (c) scope: at HEAD forecast.py serves RB/WR/TE from sim_world wherever it has a row (F-08), so the curve bands the imminent week only for QB, where the effect is nil under the runner's own arm (-0.20pp) and under the production rule (+0.41pp [-0.46,+1.24], n 1,938). The record's premise_check says this; its verdict text does not.

Next action

(1) File an addendum, do not rewrite: the +5.92pp is the previous-played-row arm; the production-selection figure is +1.88pp [1.57,2.17]. (2) Whether the residual 1.9pp is staleness or selection needs the intervention this review did not run (random-week prior row control). (3) Prospective: grade 2026 QB aar_forecasts hi80 exceedance once weeks 3-4 settle; that is the only position where the curve bands the imminent week, and it is where both arms say nil. (4) The rest-of-season weeks (k weeks ahead) remain a measurement no arm here covers.