In a model's generation settings, what does the temperature control?

The principle — temperature is a sampling setting. It changes how the next word is chosen, not what the model was told or how much it read.

Why the key is correct — Google states that the temperature controls the degree of randomness in token selection, and that it is used for sampling during response generation.

Why the others are wrong — output length is bounded by a separate maximum-tokens setting. Nothing here limits how much of your prompt is consulted. And instruction-following is a property of what you wrote, not of the sampling.

Remember this — if a reply ignores your instructions, that is a prompt problem. Reach for the prompt before reaching for a parameter.

Sources — Google's Prompt design strategies.