Anyone telling you they have tamed LLMs into producing 100% deterministic answers has either scoped the problem space so narrowly as to border on meaningless (e.g. "Is earth flat?" with a structured output schema of a single JSON boolean value), hasn't done robust statistical validation to actually confirm truly deterministic outputs, or both.
LLMs are fundamentally non-deterministic. Trying to use them to solve deterministic problem spaces is selecting the wrong tool for the job, and expecting them to be 100% reliable is the wrong mindset for working with them.
