The "not a ranking of intelligence" framing is worth stating explicitly, most setups drift the other way without meaning to. Our own pipeline runs an adversarial review pass before anything gets published, and the temptation is always to treat that reviewer as the senior one instead of a second set of eyes with a narrower question. The permission point is the actual teeth of it though: your judge has no publish/approve rights regardless of how confident it sounds. That's the real guardrail, not its accuracy.