Skip to content

Review Satisfaction

AI systems may assist review, but may not satisfy human review authority.

Formal invariant

AI systems MAY assist with review preparation, review summarization, comparison, diff explanation, evidence collection, and defect discovery.

AI systems MUST NOT satisfy a human review requirement.

A review requirement that names human review, code owner review, release review, security review, theory review, schema review, independent review, government review, or public-interest review is satisfied only by attributable review from an actor with the required reviewer authority.

Review assistance

AI participation MAY include:

  • reading the proposed change;
  • identifying affected surfaces;
  • suggesting review questions;
  • summarizing diffs;
  • locating evidence;
  • checking whether required evidence appears to exist;
  • drafting review notes for a human reviewer.

AI participation MUST NOT include:

  • approving the change as the required human reviewer;
  • fabricating review evidence;
  • marking a human review requirement satisfied;
  • rewriting the manifest to lower review requirements;
  • delegating human reviewer authority to itself.

Declared vs satisfied

A manifest declares review requirements. A validator or enforcement system checks whether those requirements are satisfied.

Declaration and satisfaction are distinct states.