Child-agent review loops

How specialist review roles feed the audit without becoming authority.

Specialist reviews are evidence inputs.

A subagent can find onboarding, runtime, copy, accessibility, or release-evidence defects.
The parent run still decides what to adopt, reject, or hand off against source evidence and owner boundaries.

Use

Child agents can review onboarding, runtime semantics, copy, tests, accessibility, and release evidence.

Boundary

Their reports are evidence inputs.
They do not override owner decisions, source evidence, or final audit closure.

Report shape

A useful report names:

findings
required patches
what closes
what remains
next smallest safe action
evidence boundary

Source surfaces

skills/implementaudit/references/child-agents.md and skills/implementaudit/templates/child-agent-report.md.

Reviewer report checklist

Findings ordered by severity.
Evidence paths or screenshots.
Specific patch target.
Residual risk and test gaps.
Recommended accept/reject decision.

Bounded roles

Docs auditor

Allowed evidence: visible page, source docs, screenshots, link checks.

Boundary: can recommend patches; cannot rewrite product claims beyond source evidence.

Runtime contract auditor

Allowed evidence: SKILL.md, templates, transcript contract, validators.

Boundary: can flag marker/order errors; cannot authorize completion.

Adversarial reviewer

Allowed evidence: failure paths, ambiguity, overclaim risk.

Boundary: findings still need parent-run triage and evidence.

Release/provenance reviewer

Allowed evidence: release readback, ledgers, changelog, package checks.

Boundary: must keep publication, provenance, and host-load claims separate.

Parent-run responsibility

The parent run must record what it adopted, rejected, or deferred.
Child-agent output should not leave orphan findings.

Adoption record

After review, normalize findings into the main implementation ledger:

adopted with patch evidence
rejected with reason
deferred with owner/source
handed off with blocker and next action

A reviewer report is not closure by itself.

Review invariants

Child-agent review is strongest when the parent run preserves the same proof boundaries as ordinary phase work.

Register any abnormal finding as an Andon candidate before patching around it.
Rerun release-proof or docs-proof checks when a reviewer report supersedes earlier evidence.
Use fixtures or canaries when a reviewer claim can be made executable.
Record AGENTS_UPDATE_DECISION when a durable anti-repeat rule is adopted or skipped.
Treat child-agent output as evidence, not authority.