Evaluation recommendations should survive unit-of-analysis reweighting
General lesson: when conversational evaluations contain repeated identical inputs or multiple decisions from one call, report sensitivity under at least three units—decision, unique canonical model input, and call. Give each unit equal total weight and average decision outcomes w…