After the last revision: calibrating release risk

Can the final revised answer be released under a defined selective-risk target?

Proposes calibrating the terminal answer object rather than the first draft, and tests whether revision and provenance features add useful release coverage. Includes exact synthetic diagnostics; no new model benchmark.

• Exact synthetic diagnostics show how candidate thresholds trade coverage for an upper risk bound at fixed δ = 5% and family size 3. These are worked calculations, not model performance.

• E12-A Exact synthetic calibration diagnostics Executed (diagnostic)
• E12-B Terminal release calibration study Proposed

• H12-1 Terminal evidence-and-revision features improve held-out release coverage by at least five points at a 2% selective-risk target.
• H12-2 Revision state adds information beyond task difficulty.
• H12-3 Answer binding catches content changed after scoring.
• H12-4 Pooled performance conceals predictable regressions in revised and conflicted episodes.

• Revision-drift guard. Flag synthesis that omits supported risks or adds untraced claims, rather than presenting it as consensus.
• Uncalibrated release-risk contract. Return a null risk score until version-bound, externally labelled calibration exists.

• Do terminal evidence and revision features improve useful coverage at identical generation compute?
• Do revised and conflicted episodes need separate risk calibration?

Builds on Round 11. Continued in Round 16, Round 15, Round 14, Round 13.