Reflective Evaluation
Requires the model to evaluate its own output before revising or committing to a final answer.
Problem
Default generation optimizes plausible output, not post-hoc checking, so constraint misses and factual errors pass through unchecked.
Solution
Insert an explicit self-evaluation phase (critique, verification, or confidence gate) before finalizing. Reapplying criteria in a second pass catches errors that generation mode often misses.