Glossary
Calibration
Calibration describes how closely stated confidence levels match observed correctness rates over comparable cases.
Evaluation
Calibration concerns the relationship between a predicted probability and the frequency of the outcome it predicts. If cases assigned about 80 percent probability are correct about 80 percent of the time, that part of the predictor is calibrated. The statement requires a defined event, population, and evaluation period. A model can rank cases well but assign poor probabilities, or be calibrated overall while performing badly for a consequential subgroup.
Evaluation groups predictions by confidence or uses proper scoring rules, then compares predicted and observed outcomes. Sample size and label quality matter, particularly near a threshold where few cases occur. Scores may shift when the source population, task, or evidence available to the model changes. A natural-language phrase such as 'high confidence' is not a probability unless mapped to a tested scale with a specified meaning.
For defensible decisions, calibration helps set review and abstention thresholds, but it does not justify a specific claim without its evidence. A 90 percent score still leaves error risk and says nothing about whether the cited passage supports the value. Precision and recall describe thresholded classification behavior; calibration describes probability quality. Both should be read beside the costs of mistakes and the prevalence of unresolved cases.
Example
In a Halden Therapeutics AG (fictional) review set, claims scored near 0.8 are checked to see whether roughly four in five are correct under the stated labels.
What calibration means and how it is measured
A system is calibrated when its stated confidence matches its observed accuracy: among all the predictions made with confidence 0.8, about eighty percent should be correct. The standard visual tool is the reliability diagram, which groups predictions by confidence and plots the observed accuracy of each group against its stated confidence. Summary measures include expected calibration error, which averages the gaps between confidence and accuracy over the groups, and the Brier score, a proper scoring rule that combines calibration and sharpness; the scikit-learn documentation on probability calibration shows both and the usual correction methods, and the Wikipedia article on the Brier score gives the definition.
The influential study by Guo and colleagues found that modern neural networks tend to be overconfident, and that a simple post-hoc method called temperature scaling is often effective in correcting it. For language models, Kadavath and colleagues found that larger models are reasonably well calibrated on some formats of question when asked to assess their own answers, and that calibration depends on the format and on whether the question is in the model's training distribution.
Why decisions need calibrated numbers
A confidence figure is only useful if people and processes can act on it. A threshold for automatic acceptance, a rule that low-confidence items go to review, and an estimate of the number of errors remaining all depend on the confidence meaning what it says. An uncalibrated score can rank items correctly and still be wrong about how many errors lie above a cut-off. Calibration is also local: a system calibrated on common cases may be overconfident on rare ones, so check by slice, such as source type, entity popularity and language. The article on confidence-aware truth discovery for long-tail evidence shows why this matters when evidence is thin.
Recalibrate when the inputs change. A model updated, a source added, or a new kind of document can shift the relation between score and accuracy. Keep a labelled sample from recent production data and recheck periodically. Report calibration with the sample size, since estimates from small groups are noisy.
A worked example
An extractor assigns a confidence to each statement about studies of Halden Therapeutics AG (fictional). A labelled sample of 1,000 statements shows that those scored between 0.9 and 1.0 are correct 82 percent of the time, and those between 0.7 and 0.8 are correct 61 percent of the time. The system is overconfident by about 10 points in the top bin and about 14 in the next. A rule that auto-accepts statements above 0.9 would admit eighteen errors in every hundred. Fitting a monotonic map from score to observed accuracy gives a calibrated score, and the team sets the auto-accept threshold at the calibrated 0.95, which holds about 5 percent errors on the sample. The threshold is now a statement about error rate and not an arbitrary number.
Checks a reviewer can run
Ask for a reliability diagram on recent labelled data, by slice. Ask how the thresholds in use were chosen and whether they refer to calibrated or raw scores. Ask when calibration was last checked and what triggers a recheck. Compare stated confidence with observed accuracy on a fresh sample.
Related terms and boundaries
Calibrated confidence is what lets a system practise abstention sensibly, and it is evaluated alongside precision and recall. It complements the estimates of the Dawid–Skene model and the match weights of the Fellegi–Sunter model, which are evidence but not probabilities until checked. It concerns the numbers a system outputs; it does not by itself make the underlying claims true.
Common misreadings
- Presenting confidence adjectives as calibrated probabilities.
- Assuming calibration transfers unchanged to a new population.
- Using a calibrated score as a substitute for source evidence.
Appears in
- Estimating source reliability
Assess a source for a defined property and period using checked outcomes and visible uncertainty.
- Measuring declared absence
Score supported unknowns as distinct outcomes, with search scope and reason assessed separately.
References
- Guo et al., On Calibration of Modern Neural Networks, ICML (2017).
- Guo et al. (2017), On Calibration of Modern Neural Networks, ICML.
- scikit-learn documentation, Probability calibration.
- Wikipedia, Brier score.
- Kadavath et al. (2022), Language Models (Mostly) Know What They Know.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Where does "Calibration" bear on a decision of yours?
This entry sits under evaluation. If a decision in your organization turns on it, send the entity concerned, the sources you fuse to describe it and who may read which of them. We reply with what a first record would cover.
Or write to [email protected] in your own words.