Fine-tuning.
Training on a narrow task sharpens the output distribution. The model becomes more accurate and more confident, and the second usually moves further than the first.
MKC2 results ↗A model that says it is 90% sure should be right about nine times in ten. Calibration is the measurement of that promise, and it decides what can be automated.
01 / DEFINITION
The short answer first.AI model calibration is the correspondence between a model's stated confidence and its observed accuracy. A calibrated model that reports 0.8 confidence is correct about 80% of the time at that level. Calibration is measured separately from accuracy, because a highly accurate model can still be badly calibrated.
Calibration is the property that turns a model score into an operational control. Every automation rule of the form act when confidence is above X, route to a human below it assumes the score means something consistent. If it does not, the rule silently sends the wrong items to the wrong place, and the failure is invisible because the model still looks accurate on average.
WORKED EXAMPLE
Imagine a support system that sends high-confidence product matches directly to a reviewer while putting uncertain matches into a specialist queue. A score of 0.9 is only operationally useful if the team knows how often predictions near that level are correct on comparable work. A model can rank the correct product ahead of alternatives without producing reliable probabilities. Test ordering and probability calibration separately.
Group held-out cases by predicted confidence and show the number of cases and observed correctness in each group. If nine of ten predictions near 0.9 are correct, report ten observations alongside the estimate. Do not present that small group as evidence that every future high-confidence prediction is equally safe. Inspect the actual mistake and ask whether a whole product family is missing from the evaluation.
If a calibration adjustment is fitted, reserve separate data for checking the adjusted probabilities. Using the same examples to fit the adjustment and judge the improvement makes it difficult to tell whether it transfers. Keep the unadjusted model, the fitted transform and the untouched comparison results together.
Finally, test the proposed review threshold against workload. Count the correct cases released, the incorrect cases released and the number referred to specialists. If a product rename changes the mix of incoming tickets, recheck those bands before allowing the same confidence value to drive the same routing decision.
Further reading · On Calibration of Modern Neural Networks
02 / WHY IT DEGRADES
Training on a narrow task sharpens the output distribution. The model becomes more accurate and more confident, and the second usually moves further than the first.
Preference training rewards decisive, fluent answers. Hedging is penalised by raters, so the model learns to sound certain regardless of evidence.
Calibration is fitted on one input distribution. New document types or a new intake process break it without changing the weights at all.
Asking a model to state a percentage in text produces numbers clustered on round values and only loosely related to its actual token probabilities.
The fourth point is worth stating plainly because it is now so common. A self-reported confidence in prose is a generated string, not a measurement, and it should be validated against outcomes before any threshold is built on it. Sometimes it turns out to be informative. It is never safe to assume so.
03 / MEASURED EVIDENCE
What a calibration record looks like.Privilege AI reports calibration for MKC2 with two figures: a Brier score, which combines correctness and confidence, and an expected calibration error, which isolates the confidence gap. Both are measured on held-out items alongside the model being replaced and the stock base model, scored in the same run.
Than both the stock base model
and the model it replaces.
Than both the stock base model
and the model it replaces.
Checks that carry no authority to change
an outcome, by design.
Calibration record, 21 September 2026. Figures are measured on held-out evaluation data, not on live matters. MKC2 runs in review-only mode: it reports and flags, it does not act.
The structural point in that record is the third figure. Where a check's ranking quality or calibration is not strong enough to support a decision, the check is given report-only authority or switched off entirely rather than being averaged into a headline number. Calibration is not only a property to measure; it is the input to how much authority a model is granted, which is the argument developed in AI model supervision.
04 / RECALIBRATION
Recalibration is a post-processing step applied to a trained model's scores. It does not change what the model knows; it changes what its numbers claim.
Two constraints are easy to get wrong. Recalibration must be fitted on data that is neither the training set nor the final held-out set, or the reported calibration error is optimistic. And because recalibration is monotonic, it improves calibration without improving ranking quality: a model that cannot order defective work above sound work will not be rescued by scaling its scores.
05 / QUESTIONS
Routinely. Accuracy counts how often the top answer is right; calibration asks whether the confidence attached to it is honest. A model that is right 93% of the time but says it is 99% sure on everything is accurate and unusable for routing, because nothing can be separated by its score.
Yes, because the confidence score determines review order and reviewer attention. An overconfident flag gets read as more serious than it is, and a reviewer who learns that high-confidence flags are often wrong starts discounting all of them. Calibration is what keeps a queue triageable.
Temperature scaling can work with a few hundred labelled items. Isotonic regression and per-slice calibration need enough items in every bucket or slice to estimate a rate, which in practice means thousands. Under-powered per-slice calibration is worse than none, because it fits noise.
Yes. The number depends on the number of buckets and whether buckets are equal-width or equal-count. Different choices can move it noticeably, so the bucketing scheme should be published with the figure and kept fixed across comparisons.
Whenever the input distribution can have moved — a new client, a new document source, a new intake process — and on a fixed schedule regardless. Calibration degrades under distribution shift without any change to the model, which is why it belongs in AI model monitoring rather than only in pre-release testing.
WORK WITH PRIVILEGE AI
Know what a confidence score is worth.
We check whether a model's confidence actually matches how often it is right.
Clear about the limits.
We report which checks are not reliable enough to base a decision on.
Better review decisions.
Trustworthy confidence lets your team focus attention where it is needed.
Talk with Privilege AI
Tell us what a reported confidence value means to your operators and which decision depends on it.
CONTACT PRIVILEGE AI
Tell us what a reported confidence value means to your operators and which decision depends on it.
Privilege AI measures ranking quality and calibration separately, and reports the checks that cannot carry a decision.