Review of 2026-09-23T2242Z-h241-qb2-carry-pool

Owner lane: Not recorded

Finding and verdict

WEAKENED

Review scope

SPLIT. UPHELD: the frozen B endpoint's direction. Scaling projected non-QB1 QBs' rush fields by P_exit before _reconcile makes fantasy points worse in both seasons, and worse on MSE as well as MAE, so it is not a mean-vs-median artefact; the shadow adds +0.5-0.6 RB carries a team-game of bias to a chain whose RB carry bias was ~0 (+0.05 / -0.03); the identity control is exactly 0.0 and pool conservation holds. B_2024 reproduces to the last digit. DOES NOT SURVIVE AS WORDED: (a) the A board block (leak 1.641, shadow -0.75, 'clean teams stay +2.885', falsifier FIRED) is a property of a served row set that no longer exists and cannot be re-derived: on the later snapshot the board leak is 0.513 (not 1.641), the shadow delta is -0.06 [-0.32,+0.23], 23 of 32 teams are 'clean' with gap -0.24, and the falsifier has nothing to fire on. (b) the mechanism sentence 'projected 0.66/0.58 vs 0.77/0.66 ACTUAL, i.e. slightly UNDER' is definition-dependent: non-QB1 is chosen by PROJECTED attempts, and in 10.7%/11.2% of team-games the projected QB1 is not the actual starter, so the 'actual non-starter' carries include those starters' own carries (0.40/0.44 per team-game). Against TRUE backups (actual starter excluded) the chain projects 0.525 vs 0.414 actual (2025) and 0.478 vs 0.285 (2024): OVER by 0.11 and 0.19 a team-game, not under. (c) 'do not scale backup QB carries in lines_for' generalises from one extreme scaling (x0.035 on the whole mass) and is untested for a smaller, identification-aware one. (d) the 2024 MAE harm does not survive the disclosed correction.