How should LLM evaluation be designed for production?

Maintain representative test sets, define task-specific metrics, combine automated and human review, track regressions by slice, and evaluate both quality and operational metrics.