DOMAIN AGENT ARCHITECTURE REVIEW CHECKLIST

Domain:
System:
Owner:
Reviewer:
Date:

1. Decision Rights

[ ] Agent decisions, recommendations, and forbidden decisions are explicit.
[ ] Human or professional review requirements are named.
[ ] Irreversible, regulated, or liability-creating actions require approval.
[ ] Escalation path is defined for low confidence or conflicting evidence.

Evidence:

2. Evidence Policy

[ ] Authoritative sources are named.
[ ] Forbidden or insufficient sources are named.
[ ] Freshness rules are defined.
[ ] Citation or explanation format is defined.
[ ] Missing, stale, or conflicting evidence has a safe outcome.

Evidence:

3. Data and Privacy

[ ] Sensitive data classes are named.
[ ] Tenant, jurisdiction, role, and user boundaries are enforced.
[ ] Retention, redaction, and deletion rules are defined.
[ ] Domain-specific records are auditable.

Evidence:

4. Tools and Actions

[ ] Domain tools have typed schemas and risk classes.
[ ] Tool permissions match role and task.
[ ] Side effects are idempotent, approved, or compensatable.
[ ] Domain policy runs outside the prompt.

Evidence:

5. Review Interface

[ ] Reviewers see evidence, extracted fields, policy checks, uncertainty, and proposed action.
[ ] Reviewer override is captured as data.
[ ] Review decisions become eval or policy updates when appropriate.

Evidence:

6. Domain Evals

[ ] Evals cover domain terminology and evidence quality.
[ ] Evals cover missing evidence, policy exceptions, privacy, and escalation.
[ ] Domain experts review labels where judgment matters.
[ ] Release gates measure domain correctness, not only prose quality.

Evidence:

7. Final Decision

[ ] Prototype only
[ ] Internal pilot
[ ] Human-reviewed production candidate
[ ] Not ready

Blocking gaps:

Next actions: