2 things.

1. If you turn the temperature down too far, the output is just bad and no amount of running prompts optimization will let you hill climb your way to good performance.

2. It’s not about determinism vs non-determinism. It’s about chaos. A perfectly deterministic model is still chaotic. Meaning that very small changes to the input result in very large changes to the output.

Turning temperature down doesn’t actually get you predictable or reproducible behavior across different inputs.