We pay people to say 'insufficient evidence': human gold-standard judges for AI grader calibration