<p>Accurate self-assessment, or calibration, is a key component of students’ metacognitive monitoring and self-regulated learning. This study examined calibration and prediction confidence across two high-stakes course exams (initial and make-up) using both continuous and group-based analyses. Undergraduate students in early childhood education completed Exam 1 (<i>n</i> = 128), predicting their grade (0–10) and reporting confidence (1–5). Students who failed (<i>n</i> = 45) took a make-up exam one month later and repeated their self-assessment. In Exam 1, calibration followed a pattern consistent with the Dunning–Kruger Effect (DKE): low-performing students overestimated their performance, mid-range students were most accurate, and high-performing students slightly underestimated their results. Among students who took the make-up exam, absolute calibration accuracy improved and confidence increased, and these changes remained robust after controlling for regression-to-the-mean effects. Together, the findings show that performance-level calibration patterns consistent with the DKE persist under authentic assessment conditions, while absolute miscalibration and confidence levels can shift following feedback and repeated testing. These results highlight adaptive shifts in metacognitive monitoring accuracy across repeated high-stakes assessments and provide practical implications for supporting self-assessment in higher education.</p>