An automatic judge rates responses, but its scores disagree with expert reviewers on consequential cases. Which approach best fits?

1. Evaluate tool selection, arguments, execution outcome, and final response.
2. Calibrate the judge against human-reviewed examples and define escalation.
3. Create a versioned golden evaluation set with representative edge cases.
4. Version components and compare controlled changes using consistent evaluations.