Skip to content
AletheIA Docs
Esc
navigateopen⌘Jpreview
On this page

Human-Review Criteria

Goal

Consolidate, in one place, when human review is mandatory before a knowledge-aware skill produces a final deliverable. The conditions are scattered across the restricted-use policy, the precedence policy, and the audit-log spec; this document is the single checklist the governance skills point to.

When any condition below is true, the skill must pause, surface a structured review request, and withhold the final deliverable until review returns.


Mandatory review conditions

Review is required when at least one holds:

  1. Mandatory-source conflict. Two mandatory sources collide on the same decision point and precedence cannot settle it (same tier, no tie-break). See source-precedence-policy.

  2. Regulatory risk. A regulated source is used in a decision with legal, financial, or contractual impact — or any source touches a named regulation.

  3. High-impact decision. The decision is irreversible, externally binding, or high blast-radius (pricing, legal commitment, public claim, safety-relevant).

  4. External publication. Output leaves the organization as a public artifact.

  5. Client delivery. Output is delivered to a specific client or third party.

  6. Sensitivity uncertainty. The resolver cannot determine a source’s sensitivity with confidence (default: escalate one level and review).

  7. Trust-boundary crossing. The task combines sources, or moves content, across project / client / organization boundaries.

  8. Source manifest opt-in. The source’s human_review_required_for list matches the current task.


What a review request must contain

A pause without context is not a review request. Include:

  • task_id, agent_id, skill_id
  • the triggering condition(s) from the list above
  • source id(s) + version(s), sensitivity, authority
  • the decision or deliverable being withheld
  • the specific question the reviewer must answer
  • (for conflicts) prevailing vs. suppressed sources

This mirrors the human_review_required / human_review_reason fields in the knowledge-audit-log-spec; the request and the log entry carry the same reason.


Behavior while review is pending

  • No final deliverable is produced.
  • The pending state travels with any handoff (see logs-and-handoffs-policy).
  • Restrictions remain in force; the pause does not relax them.
  • The request and its outcome are recorded in the audit log.

What this is not

This consolidates when review is mandatory. It does not define a review workflow, approver hierarchy, or SLA — those are project-extension and future-integration concerns, not framework core.


See also

Was this page helpful?