Output Distribution

Output distribution describes the full landscape of what a model produces — not just its most typical response but the spread of all the responses it might give to a given input. Because language models are probabilistic, running the same prompt multiple times can produce meaningfully different outputs. Understanding the output distribution tells you things like: how often does the model produce the ideal response? How often does it hallucinate? Under what conditions does it drift toward more cautious or more permissive behavior? For behavior architects, understanding output distributions is important for setting realistic expectations and for evaluating consistency — a model that produces the right answer 70% of the time may be acceptable for some applications but completely unacceptable for others.