Source: MIZOKI audit 2026-09-06 (pinned fc8b03f9) → work-order backlog 2026-09-08 (docs/audits/AUDIT_WORK_ORDERS_2026-09-08.md). Findings are hypotheses until reproduced on current main (WO-00).

Finding: Policy evaluated on data including its training rows; interval from dispersion of predicted individual effects omits fitting uncertainty [R12]. Files: cell 26 module (resolve via config/actual_urls.py in WO-00); reuse seeded-bootstrap / grouped cross-fitting from services/lift-engine. Fix: grouped cross-fit evaluation; bootstrap over refits; report exposure counts. Acceptance: leakage test: evaluation rows disjoint from training rows; interval widens when n shrinks. Depends on: WO-10.