Experts reviewing the same agent answers disagree because "good" is undefined. What should the team do to make feedback useful for iterative evaluation?

1. Discard every answer on which reviewers disagree.
2. Replace all feedback with the character count of each answer.
3. Accept whichever expert submits a score first as the permanent truth.
4. Define observable criteria, calibrate reviewers, and retain disagreement for analysis.