Your question is LLM Evaluation Under Operational Risk. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
What metrics would you prioritize for LLM evaluation in a domain where hallucinations could have significant operational consequences at Anduril?
Explain how you would define each metric, build an evaluation set, establish acceptance thresholds, and balance safety against usefulness, latency, and cost. Distinguish metrics that should block deployment from those used for continuous monitoring, and address prompt injection and inappropriate model confidence.