non-determinism

LLM Output Variance: When a Single Targeted Retry Pays Off

45% of LLM outputs changed on identical input. Here is how to turn that variance into a one-shot retry that recovers quality without burning retry budgets.