Do not ship by vibe
Many LLM features are approved because a few examples look good in a demo. That is sampling, not evaluation.
Build a test set
Collect real prompts, tickets, documents, product use cases, and hard examples. Include normal cases, ambiguous cases, malicious inputs, missing context, and examples that used to fail.
Define the rubric
Score the output on the criteria that matter for the task: factual accuracy, completeness, citation quality, format, tone, safety, latency, and cost.
Compare changes before shipping
Every prompt change, model switch, retrieval change, or data update can improve one case and break another. Regression checks catch that before release.
Keep evals after launch
Production logs should feed new examples into the test set so the eval suite becomes smarter as real usage grows.