03 of 6 · Requirement violation
The author-reviewer
The same model writes the work and reviews the work. It will approve itself roughly always.
Asking a model to objectively review its own output is asking it to find its own reasoning unconvincing — sometimes it will, reliably it won’t. The failure compounds quietly: self-review passes inflate every downstream metric, so the dashboard improves as the oversight decays. Structural separation isn’t a courtesy; it’s the only version of review that survives scale.
Review and authorship sharing an execution context, a session, or a role; verifications performed by the role being verified; approval rates near 100% with no structural reason; graders whose inputs are the gradee’s own self-report.
Separation of powers, enforced structurally — the reviewing role is a different role than the acting role, with its own charter, and the substrate enforces it, not politeness.
The standard’s verification-independence rule: the resolver that confirms an outcome runs as a role other than the one being verified — the delivery mirror of the grading-independence rule. Roles are structural (charters, gates), not prompt-flavored personas.
Official reference: Separation of powers / verification independence (Part 6 §6.6, Part 5) — the standard’s normative catalog defines ten anti-patterns; this page is the plain-English door into it.
Author-reviewer — FAQ
Isn’t using a second LLM as reviewer just the same model with a different prompt?
If both share context and incentives, largely yes — which is why the standard’s rule is structural, not cosmetic: the verifying role has a separate charter, sees the record rather than the narrative, and checks artifacts (does the row exist, does the PR merge, does the check pass) instead of judging prose. The strongest reviewers are deterministic.
Where does human review fit?
At the top of the ladder, permanently: the standard keeps the highest-consequence actions human-approved no matter how much autonomy an agent has earned. Structural separation among agents is what makes the human’s queue small enough to actually review.
Does your fleet have this one?
The free Reality Check probes for this pattern against your own record — read-only, aggregate-only, no signup to read your result.
Run the Reality Check →