Confidence Calibration Testing
Confidence calibration testing checks whether a model's stated confidence (or its output probability) actually matches its real accuracy — a well-calibrated model that says it's 90% confident should be right about 90% of the time, not 60% or 99%. Miscalibration means the confidence score is decorative rather than trustworthy.
A model can be highly accurate overall and still badly miscalibrated — it might be overconfident on the exact cases it gets wrong, which is worse than being uniformly less confident, because downstream systems (and users) that rely on that score to decide when to trust an answer or escalate to a human get systematically misled at precisely the moments it matters most.
Testing it means binning predictions by their stated confidence and checking the actual accuracy within each bin — often visualized as a reliability diagram, where a perfectly calibrated model traces a straight diagonal line and any bulge above or below it shows over- or under-confidence at that range. A common calibration metric is Expected Calibration Error (ECE), which averages the gap between confidence and accuracy across bins.
This matters most for any system that uses the confidence score to make a decision — auto-approving above a threshold, routing low-confidence cases to a human reviewer — since a miscalibrated threshold either lets bad output through or floods reviewers with cases that didn't need review.
Example
ECE = Σ (|bin_i| / N) × |accuracy(bin_i) − confidence(bin_i)|Expected Calibration Error — lower is better calibrated
