A reviewer's extra check must name what it protects

Press Release

A reviewer's request for more validation or more complexity has to say what it protects. Every abcd reviewer brief carries one rule: a finding that asks for new validation or complexity cites the acceptance criterion, the recorded invariant or the concrete risk it serves. A finding that cites none is recorded as a nit and does not force a fix.

"The review asked for a third layer of input checks," said Carol, a tech lead. "It couldn't name a criterion or an attack it stopped, so it went down as a nit, and the change stayed small."

Why This Matters

Survived, reframed, from an ideate run of 2026-10-14 (private verdict review-against-intent) on the engineering write-up "Redesigning source control at Anthropic" (Anthropic, 2026-10-13, https://www.anthropic.com/engineering/redesigning-source-control), which reports that "one such review was overly defensive about trusting another internal service, which resulted in Claude adding excessive validation and complexity". Both reviewers of the run kept line-level review (abcd's own review calibration shows small-delta reads finding defects broad passes missed) and both converged on this rule. It applies abcd's "less, but better" principle to reviews: an addition that only guards another addition is a third copy of the problem. The security reviewer's "theoretical unease is not a finding" is the precedent for one lens.

Mechanism

We expect smaller, equally safe changes because a reviewer that must cite what a check protects either finds a real criterion, invariant or risk (and the fix stands) or cannot (and the check was guarding nothing).

Scope Conditions

• Applies to abcd's own reviewer briefs (ruthless, security, adversarial ideate); line-level reading is unchanged.
• Judging whether a citation really supports a finding stays the reviewer's and the person's call; no lint decides it.

Acceptance Criteria

• Given a reviewer finding that asks for new validation or complexity and cites no criterion, invariant or concrete risk, when the review is recorded, then it is recorded as a nit and does not block.
• Given such a finding that cites one, when the review is recorded, then it keeps its severity and the citation is kept with it.

Open Questions

• Whether "a nit" or "rejected" is the right outcome for an uncited request (the two reviewers differed; product thinker at planning).

Audit Notes

Empty. Populated by intent-auditor when intent moves to shipped/.