Why Lowering Temperature Broke Daniel's Scripts

Daniel dropped his DeepSeek 4.1 scriptwriting agent to temperature 0.8, expecting tighter, more lore-faithful dialogue. Instead he got barely coherent scripts. This episode digs into the research on temperature fragility — why nearly half of tested open-weight models lose 17 to 38 accuracy points over a modest temperature shift, while the rest barely flinch — and why the failure mode is output collapse rather than wrong answers. Then it gets weirder: DeepSeek's thinking mode may silently ignore temperature entirely, meaning the 0.8 might have done nothing at all. We walk through the two-by-two diagnostic that settles it, what DeepSeek's own presets actually recommend, and why the honest answer to "what temperature should I use" is a measurement, not a number. Episode #704712 — open it directly at myweirdprompts.com/704712