input = text / json text output = answers simple questions of three flavors - score based on rubric 0-100 - yes no with confidence score - pick item out of a list of options extremely low latency compared to llm structured outputs. answers many questions against same input in parallel