Glossary

Calibration

A model is calibrated if, among predictions it labels 80% likely, about 80% are correct. Calibration matters more than raw accuracy for decision support, because reviewers need to know when to trust a score.

Why it matters

How closely predicted confidence matches observed accuracy. Understanding the term is the first step; applying it consistently is the harder part.

Browse the glossary or read the editorial standards.

Related terms

Human-in-the-loop

A design pattern in which a person reviews, approves, or overrides an automated output before it has effect.

Read more

LLM-as-Judge

Using a large language model to score or compare outputs, often as a substitute for human evaluation.

Read more