14T1420Z-premise-h648-turnover-hazard
Owner lane: Not recorded
Prediction
corr(issued, x) near 0 (team-blind) per H-648
Finding and verdict
H-648's premise as stated is not supported: the kernel issues giveaways in the same ORDER as as-of ball security (corr 0.173 [0.120, 0.228]; slope 0.060 [0.041, 0.080] giveaways per unit x), not backwards; the week-6 -0.32 was n=28 noise. Whether the slope is too SHALLOW is unresolved: real slope 0.150 [0.053, 0.253] (corr 0.085), ratio 0.40, gap +0.090 [-0.010, +0.195]; frozen call INCONCLUSIVE (issued upper 0.080 above half the real slope 0.075; lower 0.041 below it). The kernel's spread of issued giveaways ex-Monte-Carlo is sd 0.091 vs sd 0.096 the real slope supports, so total predictable spread is about right. Level: kernel 1.223 vs real 1.176 giveaways per team-game (+0.047, 4%; 2024 +0.038, 2025 +0.055), which is the larger and cleaner defect, about +0.1 points per team-game.
Next action
(1) Level audit of kernel interceptions vs lost fumbles separately against real 2024-25 (kernel lost_fumbles node exists; add INT per dropback) to name which kind carries the +0.047. (2) Slope question needs ~4x the team-games (2016-25 E-01 origins) before an as-of passer hazard is worth building; route to H-667 owner as a bundled special-teams/turnover as-of sidecar only if the level audit shows a kind-specific excess. (3) 2026 weeks 1-10 are a held-out look once 128-trajectory 2026 rollout is affordable.
Source provenance and publication scope
Owned research record: research/scientist/experiments/2026-11-14T1420Z-premise-h648-turnover-hazard.json
This public reading view includes authored question, finding, review, reasoning and next action fields. Raw measurements, commands, logs and local paths are withheld. The record ID clock is not proof of completion or deployment.