verification-audit

name	verification-audit
description	Produce, validate, and reduce the canonical Verification Record for a frozen candidate without fabricating evidence or authorization.

Verification Audit

Owns: Verification Record production, validation, and deterministic claim reduction. It is the only skill that interprets evidence, invariants, boundary deltas, gates, provider capabilities, and merge authorization into a final disposition.

The complete normative contract is:

• Verification Record reference
• Versioned JSON contract
• scripts/verification_contract.py

Other skills may collect facts or flag gaps. They must not independently implement this policy.

Inputs

Require the runner-provided normalized ticket ID, Ticket Envelope artifact reference, and frozen CandidateRef plus:

• acceptance criteria from the already-normalized ticket handoff;
• stage results for implementation, simplification, review, and QA;
• evidence records with actual class/result and causal references;
• invariant register and External Boundary Delta;
• gates and normalized provider records;
• requested operation and requested claims;
• exact-SHA merge authorization when merge is requested.

Treat absent, stale, contradictory, or unvalidated material input as unknown or blocking. Never infer live access, provider support, approval, merge success, or production behavior.

Volatile intake bound

• max_volatile_bytes: 12017 normalized UTF-8 bytes per invocation, the largest observed serialized verification_inputs payload in compact TK-01 and TK-02 checkpoint events.
• max_single_output_bytes: 12017, from that same observed payload.

Count pasted stage results, evidence bodies, provider output, and tool results after CRLF or lone-CR normalization to LF. Consume the normalized input manifest first; truncate command output before it enters context. Prefer path plus SHA-256 references over pasted artifacts and load an artifact body only when a specific causal or contract question requires it, charging those bytes to the same total. If the next required read would exceed a cap, emit a schema-3 partial checkpoint with exact remaining references and budget-exhausted; leave the fact unknown or gated. Never weaken evidence classification, causal scope, boundary accounting, verification duties, or the deterministic claim ceiling to fit the bound.

Audit flow

1. Bind every stage result and nested artifact to the same current CandidateRef.
2. Validate IDs, references, required mappings, boundary completeness, and gate consistency against contract version 1.
3. Confirm evidence crosses the changed causal mechanism and carries no stronger class than its observation permits.
4. Confirm every externally meaningful change has an explicit disposition, invariant, evidence/QA mapping, claim mapping, and gate where unresolved.
5. Normalize provider capability and mutation observations. A required unavailable capability needs exactly one explicit provider-capability gate.
6. Validate exact-head merge authorization when applicable.
7. For each schema-3 leaf, copy its isolation into existing stage limitations without changing the Verification Record schema. An independence claim with shared-context or unknown isolation requires a limitation and an explicit unsupported-independence gate.
8. Run the deterministic reducer. The declared implementation status, maximum claim, release status, and final disposition must exactly equal the reduction.

When ticket-autopilot supplies its verification checkpoint adapter, produce the semantic bundle inputs once and let the runner serialize, hash, validate, reduce, and resume them. This skill still owns every evidence classification, boundary authorization, contradiction, uncertainty, gate, and final wording. The runner must inject this skill's validate_bundle and reduce_claims functions; a checkpoint cache hit may reuse their exact prior outputs only when CandidateRef and normalized input hash are unchanged. Skipped, simulated, blocked, or unavailable live boundaries remain explicit inputs and cannot be upgraded by checkpointing.

Commands

VERIFICATION_AUDIT_ROOT means the absolute skill root resolved from the available skill catalog or from this SKILL.md location, never from repository cwd.

python3 -B "$VERIFICATION_AUDIT_ROOT/scripts/verification_contract.py" validate <bundle.json>
python3 -B "$VERIFICATION_AUDIT_ROOT/scripts/verification_contract.py" reduce <bundle.json>
python3 -B "$VERIFICATION_AUDIT_ROOT/scripts/verification_contract.py" \
  validate-pr <bundle.json> <body.md> --pr-head-sha <observed-sha>

Ticket Markdown parsing belongs to ticket-autopilot. This skill consumes only the runner-provided identity, Ticket Envelope artifact reference, and CandidateRef; it never rediscovers ticket identity from Markdown or prose.

Output

For a runner handoff, return a schema-3 verify leaf result plus the validated bundle. Bind its schema-1 quality evidence to the exact handoff-ready checkpoint hash and CandidateRef; interruptions remain non-passing partial handoffs. For standalone use, return the validated bundle and a concise audit summary:

• CandidateRef and artifact version;
• implementation status and final disposition;
• claim ceiling and permitted/forbidden wording;
• blocking gaps and open gates;
• provider limitations and merge-authorization state;
• residual uncertainty.

If no material runtime or release claim exists, say the audit is not applicable. If the contract cannot be established, fail closed rather than emitting a partial PASS.