Self-correction

Asking a model to check its own answer usually produces agreement, because the thing that caused the error is still in the context. Independence is the property that makes a second look worth anything.

Critique then revise is the simplest self-correction loop and the easiest to fool yourself with. Build it, then build the thing that tells you whether it improved anything.

Break the answer into claims, check each one on its own without the argument that produced it, and reassemble. The independence comes from what you leave out of the checking prompt.

Sample the same question several times and take the answer that comes back most often. The votes are the point, but the disagreement is the part worth keeping.

The pipeline you built always retrieves and always trusts what it gets back. Grading the results, re-querying when they miss, and deciding whether to retrieve at all is self-correction applied to the one part that can go and find the missing fact.