Calibrate and risk-tier independent artifact review

[jeremymanning]

Origin

Deferred from the MVP proposed as a fresh realization of #485.

Goal

Evolve the MVP's bounded plan/final-output review into a calibrated, risk-tiered review service.

Scope

• Per reviewer model/prompt/version precision, recall, false-positive rate, abstention, and cost metrics.
• Seeded-defect and adjudicated production datasets for calibration.
• Risk tiers: deterministic checks, sampling/differential review, mandatory independent review, and human adjudication.
• Model-family versus evidence-channel independence experiments.
• Dispute handling and accepted-risk policy.
• Review of intermediate artifacts where measured benefit justifies cost.

Acceptance criteria

• Reviewers are evaluated against a versioned benchmark and production adjudications.
• Policy selects review depth from consequence/risk and measured reviewer quality.
• Review-loop non-convergence and cost stay within declared bounds.
• A pass reports residual risk; it never claims that an artifact is simply "clean."