Review of 2026-09-23T1958Z-h114-qb-regime-judgement

Owner lane: Not recorded

Finding and verdict

WEAKENED

Review scope

SPLIT. UPHELD: every number the runner emits reproduces (64 rows, identical key set, summary identical, only row order and 10 float fields at <=1.8e-15 differ, so not byte-identical because the runner's GROUP BY has no ORDER BY); classes sum to 61; J1-J4 all have offense_pct>=0.5 and J0 none (36 started / 25 not); the 15 J1 men are all incumbent==self; the prediction was scored WRONG honestly. DOES NOT SURVIVE AS WORDED: (a) the registered falsifier verdict '40 of 61 regime QB misses carry no signal of a change' is an ARM-UNION artefact: the 61 entities are the union of four arms' regime labels (168 arm-rows; the record says 171); per arm J0+J1+J5 is S 11/29 (38%, falsifier does NOT fire; J0 = 2), SB 31/46 (67%), SBQ 31/46 (67%, the AAR default arm, fires), SQ 30/47 (64%); J0 is 23 of 25 in 2025 and 2 in 2024; (b) the post-hoc 'a knowable signal the model does not read' (J1, 8.54 vs 19.17) is conditioned on having started, and the unselected same-signal population shows the model already prices it as a mixture (see 3_METRIC/4_SELECTION); (c) 'regime' in attrib.py is the largest of four decomposition components on an out-of-80%-interval miss, not 'change of starter'; the record's 'regime QB errors' is a population of largest-component misses and its J0 class (over-projected non-starters, the biggest) is a stack/calibration term, not a missed start.