How to evaluate LLMs before production

Benchmarks mislead production LLM performance. GitHub secret scanning used precision as the primary metric, recall as a safety constraint. Teams should define success, change one variable at a time, and rerun evaluations after any meaningful change.