The model that does the work never reviews the work. That was one of the first rules I wrote for myself, and it's the one I've broken most often.
I'm building a multi-agent AI system in my own time, mostly to learn how far I can trust one. I thought writing the rule down was the hard part. It wasn't.
One piece of work passed four rounds of review, approved every time, by models from the same provider that produced it. When I finally put it in front of a different provider, it came back with 24 more problems. Not typos. Blockers. Six more rounds to clear them.
Another time I waived the rule on purpose. Quota was tight, the change looked small, so I let it ship on a same-provider review. The independent pass ran later anyway and found eight reproducible issues, four of them serious. Nothing caught fire. All of it had been sitting there, approved.
The one that changed my thinking was a review that reversed me. I'd evaluated a set of open-source coding tools and picked a winner. The reviewer ignored my summary and went back to the raw transcripts. My top candidates had already failed an isolation test I'd recorded as passed. The models inside them had resisted the planted instructions. The tooling around them hadn't. I withdrew the recommendation, and the answer became “none of these yet.”
Agreement isn't evidence. Independence has to be structural, not polite.
Here's the part that's specific to AI. Two models trained on similar data can share a blind spot and agree with complete confidence.
So what I check now isn't who reviewed something. It's whether the review could have changed the answer. Did the reviewer get the real evidence, or my summary of it? Was “this is wrong” a sentence they were allowed to write? A reviewer that can only approve is decoration.
Every review now ends with a record of what I accepted, what I challenged, and what I deferred, with the reasoning left visible.
None of this proves a second provider always finds more. Those later reviewers also saw more finished work and asked different questions. And I don't cross-review typo fixes.
I'm still working out where the line sits.