Validate

Validate

Independently judge one exact subject against the acceptance in its existing bead or caller source, return one semantic result, and stop. Validate is the sole semantic author of verdict.v2 when persistence is requested. ao provenance store-verdict supplies structural verification and atomic storage. Before the verdict, read boundaries.md in the rpi skill's references directory for the state Validate leaves to the caller.

Prompt

Validate bead ag-1234 in this fresh context. Intent: the bead text and digest.
Subject: manifest.json from:
ao provenance manifest --root . --include cli/internal/gates
Author context ctx-a1. Re-run `cd cli && go test ./internal/gates/...`.
Return PASS, FAIL, or NOT_PROVEN with evidence; stop.

Preconditions

• The subject is a nonempty implementation candidate: the manifest lists at least one entry. Plans, audits, and reviews are subjects only when the caller explicitly requested document review.
• The intent source is a caller-owned artifact or a runtime-owned content-addressed snapshot; its acceptance digest is derived automatically.
• Author and validator context IDs are explicit, and freshness is attested with source: runtime | caller and an attester identity. Missing, colliding, or unattested identities produce NOT_PROVEN: a declared trust fact, not cryptographic proof of isolation.

Fresh validator and model selection

Default to one fresh, author-distinct validator from the author's model family: Codex/OpenAI work uses a fresh Codex/OpenAI reviewer; Claude/Anthropic work uses a fresh Claude/Anthropic reviewer. Use the runtime's configured capable model unless the caller pins one. Fresh context is required even when model weights are identical; a new role instruction in the author's session is not fresh. Risk determines the depth of evidence inspection, not an automatic second family.

Acceptance, tests/gates, stopping, allowances, safety, disclosure, hooks and executable controls warrant deeper checks, including policy written as prose. Conservative risk cues include cli/internal/gates/**, scripts/check-*.sh, tests/**, skills/*/scripts/**, skills/cc-hooks/policies/**, lib/**, .github/workflows/** and scripts/security-gate.sh. Unknown risk receives deeper inspection; it does not silently change the selected model families.

The caller can request --cross-model or say "cross-model review" to add one fresh validator from a different family. --cross-model <model> pins that additional reviewer, for example --cross-model claude-fable-5-1 from Codex or --cross-model gpt-6-astra from Claude. These are skill prompt options, not ao CLI flags. RPI forwards them unchanged. Without a pin, use an available, authorized capable model from the other family; never silently substitute for a pinned model or count two models in one family as cross-family diversity. An explicit caller requirement remains required until the caller changes it; changing the default cannot erase a finding or relabel a missing verdict as PASS.

Route selection and invocation through agent-native model-dispatch; references/mechanics.md owns evidence storage. There is no fixed ten-minute review timeout. Use the caller's selected review timeout or remaining native deadline, respecting any earlier host or goal limit. A timeout is missing judgment, not FAIL; never restart to renew an allowance.

Each selected judge receives the exact subject, unchanged acceptance and authorized evidence independently, without the author's desired verdict or peer conclusions. Record actual model/context identities and runtime receipts. Missing freshness, an unbound subject, incomplete acceptance evidence, or nonempty not_checked prevents PASS regardless of model family. With no authorized adapter for requested diversity, disclose diversity_unsatisfied: the required combined result is NOT_PROVEN even if the same-family judge passed. A delivered FAIL stands. Advisory diversity that the caller explicitly made optional may accompany the same-family result with that limitation; it cannot discard an acceptance-relevant finding.

When selected judges disagree, preserve both verdicts and their evidence. Required diversity converges only when both pass; neither majority vote nor a preferred judge settles a split. Repair addresses findings under the existing convergence law. Report unresolved dissent and the caller's decision openly. Same-family fresh judgment reduces anchoring, but does not prove independence from shared training biases; cross-family agreement is corroboration, not truth.

Mutating-check quarantine

Classify every acceptance-listed command as read-only or subject-mutating before running it: regen scripts, sync scripts, formatters, and anything with --force are mutating until proven otherwise. Run a mutating check only against a disposable copy or a committed subject, never the judged working tree (the boundaries appendix records the regen that overwrote a subject).

Scope disclosure

not_checked has exactly one meaning: in-scope acceptance surface this validation did not verify. PASS asserts the whole declared acceptance surface was verified, so a PASS carries no not_checked entries; every other scope limit has a home that survives inside a PASS: a bounded proof in criteria[].reason, a declared non-goal in the intent source, residual risk in the report (table in the mechanics reference). Emptying not_checked to obtain PASS is a contract violation: unverified acceptance makes the honest result NOT_PROVEN, and an entry that was never acceptance moves to its home and stays visible. A finding necessary to acceptance cannot be relabeled as optional, residual risk, or a non-goal to obtain PASS.

Workflow

1. Derive subject-manifest.v1 with the helper's manifest command (flags in the mechanics reference) at the start and again at the end; any mismatch is subject mutation and returns NOT_PROVEN.
2. Confirm the intent-source digest is unchanged since implementation, every cited evidence digest matches the artifact it names, and complete changed-path coverage can be derived; otherwise NOT_PROVEN.
3. Adjudicate the actual diff: runtime-derived changed paths against the intent's scope classes. A proven out-of-scope path is FAIL; incomplete scope evidence is NOT_PROVEN.
4. Inspect the exact subject and evidence. Reported exit codes are claims: re-execute the proofs that bear on acceptance. A changed test, gate, fixture, golden, tolerance, suppression, or acceptance source must be required by the original intent, with green coming from implemented behavior; green obtained by weakening acceptance is FAIL. Judge every acceptance criterion against its own evidence reference; a criterion with no evidence of its own is unverified, not passed.
5. Choose exactly one semantic result: PASS, FAIL, or NOT_PROVEN. Return it with criterion-level results, findings, evidence references, checked, not_checked, both identities, both context IDs, and the freshness attestation. Name a class for each finding: one short stable name for the kind of defect, one per finding, reused word for word when the same kind recurs, so the orchestrator can see a closed kind come back. Explain in the existing summary and evidence whether a newly exposed defect pre-existed the change, was introduced by it, or has unknown cause. Use before/after proof or equivalent causal evidence under unchanged acceptance; counts and timestamps do not establish cause. Recurrence calls for causal examination and does not by itself prove a design failure. Name the class or omit it; a class that is present and blank is a finding against this validator, and so is a class that does not describe its finding. PASS requires distinct identities, explicit freshness, nonempty checked scope, nonempty top-level evidence, evidence for every criterion, and an empty not_checked. A documentation sentence claiming something is published, pinned, or proven is an acceptance criterion like any other: it needs a check this validator can run, or it is not_checked. The docs.claims-tracked gate covers the tracked-file half of that and nothing more.
6. Only when the caller requests machine-readable evidence or a declared downstream consumer requires it, persist canonical verdict.v2 with the helper's store-verdict (mechanics reference), then return the artifact path and digest with the result. Stop.

Fresh validation is independent judgment over the exact subject, not a replay of every author command: rerun the risk-critical, uncertain, or thinly evidenced checks; a digest-bound deterministic receipt may prove routine facts; replay an expensive full suite only when acceptance requires it. The repository's full literal CI command set, as quoted in AGENTS.md, runs once, on the final integrated subject.

It's working if

Observable in the trace, without reading the prose, and the rubric a fresh independent judge scores this skill against:

• A criterion whose evidence is a justification rather than a proof is named, and the result is NOT_PROVEN rather than PASS.
• Green obtained by widening a tolerance, skipping a case, or re-baselining a budget is reported as FAIL, never as completion.
• Every scope limit is placed in one of the Scope-disclosure homes; none was deleted to reach PASS.
• The subject manifest is derived twice, at the start and at the end, and the two are compared.

Boundary

Validate emits no next action, repair, retry, replan, or delivery state (full list in the boundaries reference); ledger availability cannot change a verdict's validity.