CCDV-F: Claude Certified Developer — Foundations

LLM generation is probabilistic: sampling can select different plausible tokens for identical inputs, so minor variation across runs is expected and does not by itself indicate an application defect. Production applications should manage this variability with measures such as output validation, retries, and—where supported by the selected model—sampling controls such as temperature. Anthropic notes that API outputs can remain non-deterministic even at temperature 0.