[cschanhniem]

Summary

Adds the missing permission layer between v0.5.19 calibration-intake (descriptive join) and any future recalibration engine. The most dangerous failure mode for the wet-lab compression roadmap is silent recalibration: an agent sees real wet-lab data, decides the scorer "needs improvement," and rewrites weights to fit. The new gate makes that impossible without violating a pre-registered, machine-enforced policy.

What ships

Policy contract (configs/recalibration_policy.yaml)

• 7 minimum_conditions (cohort size, controls, orphans, positives, negatives, metrics availability)
• 5 prohibited_actions permanent floor (no toxicity relaxation, no hemolysis relaxation, no novelty relaxation, no pathogen optimization, no post-hoc success redefinition)
• 2 rate_limits (L1 weight budget 0.10, cooldown 14 days)
• 3 required_reviewer_artefacts (intake JSON, intake Markdown, dated decision log entry)
• 12 locked_changes entries, one per enforced rule

Modules

• openamp_foundry.calibration.policy: load_recalibration_policy parses and validates the policy. Rejects policy files that drop canonical prohibited actions or remove locked_changes entries.
• openamp_foundry.calibration.recalibration_gate: evaluate_recalibration_gate returns a GateVerdict with may_recalibrate (bool), per-rule results, prohibited-action audit, rate-limit status, reviewer-artefact status, reasons, and summary. JSON + Markdown writers.

CLI

• openamp-foundry recalibration-gate (exit 0 = may recalibrate, 3 = forbidden, 2 = input error)
• Makefile targets: recalibration-gate-example, recalibration-gate INTAKE=... [DATE=...] [PREV=...] [L1=...]

Tests

39 new tests, total 1647 passing (was 1614). New file: tests/test_recalibration_gate.py covering policy loader (happy + every rejection mode), gate evaluator for every minimum condition, prohibited-action audit, rate-limit status (unknown/ok/exceeded), reviewer-artefact status, writers, end-to-end CLI smoke.

Docs

• docs/CALIBRATION_POLICY.md (new): human-readable overview
• CHANGELOG.md, docs/ROADMAP.md, docs/METRICS_CURRENT.md, AGENTS.md: v0.5.20 milestone
• docs/ARCHITECTURE.md, docs/PLAN.md: link policy into wet-lab compression loop and Phase 5

Honest limitations (must read before relying on the gate)

1. The gate does NOT trigger any weight update. It only emits a verdict.
2. A may_recalibrate=true verdict is a permission, not a command. The decision to apply a weight change still belongs to a human reviewer with a dated decision log entry.
3. The gate evaluates cohort evidence, not pipeline calibration health. Benchmark regressions are caught by make validate-scoring, make bench-triage, and the selectivity benchmark — not by this policy.
4. The synthetic example correctly yields may_recalibrate=false (cohort size 4 < 5, one positive control failed, all reviewer artefacts missing). That is the expected outcome on tiny synthetic data and is itself a useful sanity check that the gate is enforcing the cohort floor.
5. The five canonical prohibited actions are duplicated from AGENTS.md and MISSION.md. The validator rejects a policy file that drops any of them. Any relaxation of the source documents must happen in lockstep with the policy file.

Verification

1647 passed, 7 skipped in 31.84s
ruff check src/openamp_foundry/calibration src/openamp_foundry/cli tests/test_recalibration_gate.py → All checks passed!

Co-authored-by: OmX omx@openamp.local