Recall
What is recall?
Recall measures how well a model or classifier performs on positive cases—the cases where the thing it is looking for is actually present. In machine learning terms, it's the proportion of true positives the model correctly catches. A model with high recall misses very little. A model with low recall lets real cases slip through undetected—these misses are false negatives.
How does recall apply to LLM-as-a-Judge evals?
When you calibrate an LLM-as-a-Judge against a set of samples you've manually labeled, recall measures the judge's accuracy when there is an error. When a sample truly contains the error you're hunting for, how often does the judge catch it? It's one of two measures to check before trusting a judge—the other is specificity, the judge's accuracy when there is no error.
Why doesn't perfect recall prove a judge is trustworthy?
Perfect recall only tells you the judge catches every real error. It says nothing about how often the judge flags errors that are not there. For example, one judge built to catch missed sub-groupings in AI-generated opportunity solution trees scored 100% on recall—it correctly identified every instance where a missed grouping existed. But its specificity was terrible: it also identified missed groupings where there were none.
A judge like that floods your measurements with false positives, so you can't trust its error counts even though it never misses a real error. Always score a judge on both measures before you trust its judgment.
Related terms:
Learn more:
Last Updated: September 16, 2026