A team compares two models for answering customer-service questions. Answers can be phrased differently while still being correct. Which evaluation practices provide stronger evidence than comparing one attractive demo response? Select 2 answers.

1. Pick whichever model produces the longest response to one easy question.
2. Include qualified human review for nuanced or high-impact judgments and track disagreement or error patterns.
3. Use different test questions and different quality criteria for each model, then compare only their totals.
4. Evaluate only whether the endpoint returns HTTP 200.
5. Use a representative held-out question set and a consistent rubric for correctness, completeness, and policy compliance.