L08-05 — Calibrate task and stratum abstention #65

Outcome

Measure error, utility, confidence, and support per question family, per head (operation vs target), and per relevant print/customer stratum. Map these to preregistered thresholds that decide when the operation head may yield anything other than ABSTAIN.

Parent and boundary

Parent: #9. Owner: proposed Decision module. This issue owns one bounded slice of Bounded Decision Worker; consume accepted upstream contracts and keep PDF/verdict/publication authority with Core.

Acceptance

• Thresholds are preregistered per task, head, and stratum. Unsupported slices abstain with a reason.
• Calibration covers the selected head's probability, not just the reported confidence scalar. Target-head reliability is measured only on cases where its operation was selected.
• Repeated no-progress decisions within a session (jev: three unchanged steps → BLOCKED) abstain rather than retry.
• Failure case: one universal probability, an argmax that clears validation, or incomplete context cannot claim safety.

Required proof

Calibration reliability, selective-risk, and OOD test packet, split by head and stratum. Record the exact source SHA, test command or hosted run, fixture/artifact identity, and any skipped or unavailable lane. Run the repository's mapped proof and check-change gate for a PR.

Handoff

Record changed files, observed result, remaining risk, and a 1–3 sentence anti-slop review. Completion means acceptance evidence, not merely a merged implementation.

Sources

Approved master roadmap · North Star: jev-ultrafast@1231850a; see the translation table in #9.

Generated by Claude Code