LLM evaluation

LLM evaluation combines offline test sets, human review, automated graders, regression tests, production monitoring, and task-specific metrics. Strong evaluation looks beyond generic benchmark scores and measures groundedness, tool accuracy, latency, cost, refusal behavior, and failure cases for the actual user workflow.