LLM Testing: Systematically validate factual accuracy, refusal behavior, and safety boundaries

Build LLM tests for factual accuracy, citations, refusal, privacy, latency and cost; pin model settings and repeat runs to assess stochastic output.