Linguistic calibration through metacognition: aligning dialogue agent responses with expected correctness