A team uses a model-generated evaluation report as the only evidence for a high-impact release. Which two improvements are appropriate? Select 2 answers.

1. Remove every failing example before calculating the final score.
2. Treat a fluent explanation as proof that every score is correct.
3. Validate automated evaluation against relevant expert-labeled examples and a clear rubric.
4. Let the candidate model approve itself without independent criteria.
5. Include independent human review where the risk and acceptance requirements call for it.