A confidence threshold belongs to the model version

A model migration can pass its accuracy and integration checks while changing the effect of an established review policy. The replacement may assign different scores to the same cases, so an unchanged threshold can send a different workload to the review team.

A threshold such as 0.8 combines a business decision with a model-specific scale. The business decision concerns acceptable errors and review capacity. The scale concerns how that version relates its scores to outcomes. Preserving the number alone does not preserve the policy.

Validate the operating threshold for each model version. Relate the replacement’s scores to observed outcomes on the intended population, then approve the point at which cases proceed or reach a person. Retain the previous risk decision where it remains appropriate, while establishing how the new version implements it.

Accuracy does not preserve the score scale

Suppose a hypothetical classifier routes low-confidence cases to human review. On the old model, a score of 0.8 might identify a band in which about eight of every ten predictions are correct. That interpretation holds only if the model is calibrated for the population in question. It means that among predictions given 0.8 confidence, the expected proportion correct is 0.8.

Guo and colleagues use that definition in their study of neural-network calibration. They found that modern neural networks could improve classification accuracy while producing confidence estimates that overstated correctness. The architecture, training choices and regularisation affected calibration. Their practical result was that temperature scaling, fitted on a held-out validation set, often improved calibration without changing the predicted class. On Calibration of Modern Neural Networks, ICML 2017

That separation matters during replacement. Accuracy asks how often the chosen class is right. Calibration asks what a score means. Two models can have similar accuracy and different mappings from score to outcome. A model can also become more accurate and less calibrated. Passing an accuracy gate therefore says nothing about whether the old threshold still selects the intended cases.

The decision to preserve	The evidence to rebuild
Which cases may proceed without review	The new score bands and their observed error rates
How many cases the review team can absorb	The new version's score distribution on representative traffic
The relative cost of a miss and an unnecessary review	The new operating curve across plausible thresholds
Which groups require separate protection	Calibration and error measures for those groups where the data supports them

The comparison separates policy from measurement. A migration record should show which requirements remain fixed and which score relationships have been tested again. This makes any change in the review boundary an explicit decision rather than a side effect of configuration.

illustrative fixture

Does the score match the outcome?

Constructed fixture: 300 labelled cases in three bins of 100. All fixture cases are included; real coverage and subgroup results require local data.

Swipe or scroll for the full diagram →

LowMiddleHigh

Low: 20 successes / 100 cases; frequency 0.20; interval 0.13–0.29.

Bin	Stated	Successes / cases	Interval
Low	0.3	20 / 100	0.133–0.289
Middle	0.6	50 / 100	0.404–0.596
High	0.9	70 / 100	0.604–0.781

Calibration concerns agreement between probabilities and outcome frequencies in this population. It establishes neither universal accuracy nor safe transfer to another use.

Recalibration is part of migration acceptance

Start with scores from the frozen replacement candidate on data that resembles the work it will receive. Do not fit calibration on the training examples. Use held-out, labelled cases and retain enough untouched data to test the result. Plot a reliability diagram or an equivalent table that groups predictions by score and compares confidence with observed correctness. Report how many cases sit in each band, because a visually neat band supported by few observations cannot carry a consequential decision.

Then evaluate operating points. For each plausible threshold, show the errors that would pass, the correct cases that would be escalated, and the volume sent to the human queue. Where consequences differ by group, check those quantities separately rather than allowing the largest group to determine one reassuring average. The threshold owner can then choose the point that implements the existing risk and capacity decision, or decide explicitly that the decision should change.

The record explains the number that will control the live workflow. It also provides a reference against which fresh labels can reveal whether the score-to-outcome relationship is changing. A benchmark result alone cannot supply either function.

A calibrated threshold can still expire

Calibration on a held-out set is conditional on the population represented by that set. Guo and colleagues state that their calibration methods assume the training, validation and test data come from the same distribution. Production rarely grants that assumption indefinitely. On Calibration of Modern Neural Networks, paper

Ovadia and colleagues tested uncertainty methods under increasing dataset shift across image, text, categorical and genomics settings. In their experiments, accuracy and the quality of uncertainty generally degraded as shift increased. Temperature scaling fitted on independent and identically distributed validation data worked under that condition and small shifts, then lost ground as shift grew. The study does not provide a universal expiry period. It shows why a calibration result from migration day cannot be treated as permanent. Can You Trust Your Model’s Uncertainty?, NeurIPS 2019

The production design therefore needs current labels, not just score and input drift. Monitor the relationship between score bands and outcomes on the traffic the system actually receives. A changed score distribution can tell the team where to look. Only observed outcomes can show whether the confidence attached to those scores still deserves its interpretation. This is the migration consequence of the institute’s earlier distinction between a drift alert and a performance alarm.

The NIST AI Risk Management Framework places this inside a continuing measure function. It calls for pre-deployment and regular operational testing, measures of uncertainty and documented results, with responses that can include recalibration, mitigation or removal. The framework is voluntary and does not prescribe a calibration method or threshold. It does support the operating principle that uncertainty evidence is maintained across the lifecycle rather than accepted once. NIST AI RMF Core, Measure

What this does not tell you

Calibration does not make a model safe, fair or suitable for a particular use. A well-calibrated model can still have unacceptable error rates, omit a subgroup or optimise the wrong target. A single aggregate calibration measure can also hide serious local errors. The method must be chosen for the decision and the available data.

Nor is temperature scaling a universal remedy. The cited results cover specific models and datasets, and the evidence under shift shows its limit. Some systems produce scores that are not intended to represent probabilities at all. In that case, the team should call them scores, test how they rank cases and avoid presenting them as confidence until an empirical mapping supports that word.

The earlier article on model retirement argues for a runnable evaluation set before the supplier’s deadline. Add the review boundary to that acceptance process. The model owner should be able to identify the cases that proceed, the cases that reach a person and the outcome evidence supporting the chosen threshold.