Your human evals need calibration too. @cwolferesearch explains the loop: define quality, have multiple people rate the same examples, then use their disagreements to improve the guidelines. #AIEngineerNY