Safety Score Curation

Trigger And Exclusions

Use for active-model curation and score-affecting refreshes, not other scoring models or copy-only edits. Stable safety-score-v9 paths/commands do not identify the current methodology version; read shared/lib/methodology-versions/current-version.json.

Classify The Operation

Select evidence curation, weekly expiry sweep, mechanism/shock/protocol producer refresh, DDRR calibration, equivalence testing, or methodology change before loading branch context. Never infer evidence from a schema gap.

Mandatory Core

• shared/data/safety-score-v9/AGENTS.md; Active model, Canonical Publication, and equivalence gate selection.
• Inspect shared/data/safety-score-v9/methodology-policy-candidate-v1.json for policy and scripts/lib/automation-registry.mjs for artifact ownership.
• Never invent, extrapolate, or date-bump evidence. Missing evidence stays bounded/open. Preserve fail-closed publication: global holds retain accepted ratings; asset-local defects may publish technical pipeline gaps only within assessV9Publication's partial-publication allowance in worker/src/lib/safety-score-v9/publication-assessment.ts.
• Attribute code and curation separately; never silence unexplained replay drift or declare unexpected movers acceptable. Never hand-edit shared/data/safety-score-v9/evaluation-build-manifest-v1.ts.

Branch Reads And Actions

• Curation/expiry: read sweep capture and typed missing-data registry. Generate/drain safety-score-v9:missing-data-registry and safety-score-v9:curation-worklist with the linked capture/output arguments, claim-group decisions, current primary evidence, and named promote/reject/defer triggers. safety-score-v9:expiry-queue is preventive, not a complete missing-evidence inventory. For the weekly sweep, also read weekly close, run safety-score-v9:live-withheld -- --replay <same-replay> --output <path>, and drain DEP/RESV, pre-expiry and counterfactual lanes; the other queues omit live-to-fallback grade drops.
• Mechanism overlay: read evidence classes and requirements. Use actual evidence pins, never capture dates. Date-only reviews enter after their UTC day elapses and temporarily re-attribute that day's admission gap method-unsupported; exact UTC timestamps enter at their instant (isMechanismOverlayCurrent in worker/src/lib/safety-score-v9/extension-mechanism.ts). Both retain original-start expiry; component-only reviews never renew siblings or an expired whole overlay.
• Shock producer: read workflow and load-bearing attestations; require evidence-backed overlays and shock attestations.
• Protocol producer: read artifact contract. USDe/USDf reviewed weekly/manual automation is permanent non-publishing evidence tooling. Strict replay needs original-byte canonical V2 or exact hash-verified normalized-only V1; summary recognition/unavailable bodies are not success. Network-free replay requires local originals/cache; otherwise signed R2 reads and verified cache writes occur.
• DDRR: read commands and guardrails; results are advisory.
• Pinned on-chain research: use npm run research:dwellir-rpc -- under pinned-evidence rules. Cite the keyless URL, block and timestamp provenance, never latest.
• Replay/equivalence (any score-affecting change): read harness capture, replay, diff, and post-deploy. Pin the capture clock; preserve captured redemption by default. --rederive-current-redemption is only for expiry/current-curation scenarios, never accepted/embedded-registry captures or --registry-ref; registry mismatch is only for attributed code/curation comparison. Require two independent captures at least 90 minutes apart, normalized score-output equality for neutral changes or prospective reviewed grade/status movers. Malformed/duplicate cards/manifests fail. Manually close absent declarations/census/same-grade findings; unexpected movers stop release. Independently verify activation identity and source-revision/registry/clock/accepted-enrichment reproduction: an empty normalized diff is not bit-identical identity proof.
• Methodology: follow ADR-3, numeric-decimal versions and all four update targets, not semver minor.

Checks Owned By The Verifier

The assigned verifier runs applicable checks after writers finish; report unrun requirements rather than claiming proof.

• Score-affecting work: npm run safety-score-v9:replay, npm run safety-score-v9:diff, npm run safety-score-v9:movers, with branch-owned capture/output arguments.
• Curation/sweep: the missing-data/worklist commands above; preventive expiry and weekly live-withheld only when applicable.
• Evaluation inputs: npm run generate:safety-score-v9-evaluation-build, then npm run check:generated-artifacts -- --only=safety-score-v9-evaluation-build.
• Replay/overlays: npx vitest run worker/scripts/__tests__/replay-safety-score-v9.test.ts worker/scripts/__tests__/diff-safety-score-v9-replays.test.ts shared/types/__tests__/safety-score-v9-overlays.test.ts.
• After authorized deployment, complete the linked first-cycle gate; local replay does not establish production acceptance.

Completion Evidence

Report work/assets, sourceGeneration, baseInputGenerationId, clockSec, evidence, replay/diff/movers, artifact ID, methodology targets when applicable, actual verifier results, post-deploy acceptance or pending evidence, and risks.