On this page
Completion semantics
What AUDIT_COMPLETE means and why run completion cannot come before it.
Done means final checkable evidence, not a stopped chat.
AUDIT_COMPLETE is the review record that every in-scope item closed with checkable evidence.
If a real blocker remains, the run emits AUDIT_HANDOFF instead, and IMPLEMENTAUDIT_RUN_COMPLETE must not appear.
Valid completion
AUDIT_COMPLETE is the final audit verdict over the live work record.
Maintainer formal name: tdqyq-audit-object.
It means:
- every in-scope item has final checkable evidence
- every claim is bounded by its evidence type
- every unresolved item is explicitly classified
Only then can IMPLEMENTAUDIT_RUN_COMPLETE appear.
Audit gaps
If the final audit finds a gap, the run writes AUDIT_GAPS, fixes with new evidence or a changed approach, and audits again.
There is no arbitrary retry cap; closure depends on evidence, not on the number of attempts.
Invalid completion
A stopped chat, failed check, hidden retry, unresolved owner decision, handoff marker, or missing evidence is not completion.
Evidence type taxonomy
Every behavioral claim should name the evidence type that supports it:
live runtimelocal generated-runtimepackage-boundunit testintegration teststatic checkermanual inspectionvisual/browserunverified
Static evidence is never upgraded to live proof.
What complete does not mean
- Marketplace verification, listing, or publication.
- Universal install proof across hosts.
- Provenance, signing, attestation, SBOM, or checksum chain.
- Graphify proof; Graphify remains orientation evidence.
- ActiveGraph correctness or lifecycle proof; it may assist fork/diff checkpoints or mirror already-recorded run-root events.
- A felt sense of completion. Evidence determines closure state.
Source surfaces
skills/implementaudit/templates/PROTOCOL.md final audit and skills/implementaudit/references/transcript-contract.md.
Marker ordering
AUDIT_COMPLETE must precede IMPLEMENTAUDIT_RUN_COMPLETE.AUDIT_HANDOFF must not appear with IMPLEMENTAUDIT_RUN_COMPLETE.
Final audit checks
Final audit should inspect the relevant evidence surfaces, not a one-size-fits-all folder checklist.
Bound work record
- When it applies
- Every run
- What reviewer checks
- Request, scope, owner/source, claims, changed files, checks, and closure state.
Run folder (run root)
- When it applies
- Phased or larger work
- What reviewer checks
- Roadmap, state, phase checklists/specs, thinking/protocol files, sidecars, and final fix-and-verify notes.
Baseline and deliverables
- When it applies
- Every mutation claim
- What reviewer checks
- Baseline reference, deliverable comparison, generated-artifact freshness, and audit coverage math.
Closure markers
- When it applies
- Every completion or handoff
- What reviewer checks
- Marker order, handoff exclusivity, unresolved risk, and sidecar boundaries.
Completion checklist
- The bound work record exists; for phased work, the run folder,
STATE.md, roadmap, and phase checklist exist for the in-scope work. - Each edited scope has before/after evidence or a named reason that evidence is impossible.
- Every failed check has a class, cause, countermeasure, rerun evidence, and outcome.
- Each phase verification includes a
Cleanliness:section tied to the baseline reference. - The final audit includes a
Deliverables:block from repo-state comparison. - The final audit includes
Audit coverage:so reviewers can see what was rerun, skipped, or trusted from prior evidence. - Generated artifacts are either verified against source or explicitly excluded from claims.
- Owner/source, scope, and authorization decisions are recorded where they affected the run.
- Any trusted earlier evidence is named; if accepted earlier evidence (
trust-prior) exceeds 30% of rerun-plus-trusted checks, the run must printAUDIT_WARNING: trust-prior > 30%instead of hiding that reliance. - Deliverables are non-empty when the request expected an artifact.
Final state checklist
All ledger items final
Every item is done, changed, blocked, deferred, or unverified; nothing is orphaned.
Evidence type is named
Each claim cites the right evidence class: file diff, command output, generated artifact readback, release readback, checksum integrity, or explicit handoff.
Generated artifacts checked
Generated output is regenerated from source or intentionally excluded from claims. Hand-edited generated output is not silently accepted.
Release and publication gates separated
Commit, push, tag, release, publication, host install, and provenance are not implied by local checks.
Marker order valid
AUDIT_VERIFY precedes AUDIT_COMPLETE; AUDIT_COMPLETE precedes IMPLEMENTAUDIT_RUN_COMPLETE.
Handoff is honest
If blocked, the run leaves AUDIT_HANDOFF or ANDON_HANDOFF, blocker, remaining risk, and smallest next owner action.
Phase outcomes
A phase can:
- complete
- defer with an owner decision
- escalate with new evidence
- hand off when blocked
It cannot disappear because the final audit found the phase inconvenient.