You’re right that once merged-with-no-review becomes normal, it stops looking like a decision. That’s why I’d track the unreviewed-merge rate explicitly. It turns a silent default back into a number someone has to answer for. On your question, I’d separate evidence from the rubric. Agents can and should generate evidence, because that’s the cheap part. What they can’t be allowed to do is define what counts as passing. In practice that means three things: A human writes the acceptance criteria before the work starts, and they sit somewhere the building agent can’t edit. The verifier runs in a different session from the builder, with different context. You check the tests themselves, for example with mutation testing: if you deliberately break the code and the tests still pass, the evidence is theatre. The risk you’re describing is real when one agent writes the code, the tests and the definition of done. Split those three apart and the human is grading against their own rubric again.
