Using one model to score another's output against a rubric, so quality can be measured at a scale human grading cannot reach.
For open-ended work there is no exact-match answer to compare against, so a judge model reads the output and scores it. It is the only way to evaluate thousands of generations, and it is how most eval suites for writing, summarization, and agent behaviour actually work.
Judges have consistent biases: they favour longer answers, prefer their own style, and are more reliable comparing two outputs than assigning an absolute score. Calibrate against human labels before trusting the number, and prefer deterministic checks wherever one is possible.