LLM-as-judge
Using a model to grade outputs against a rubric, where no code check can decide correctness. Cheap and scalable, and itself something that has to be validated against human labels.
What people get wrong
Judges have systematic biases: for longer answers, for their own family's style, and for whichever option came first. Measure agreement with human graders before trusting the score.
A definition is not an understanding. This is taught properly in Grading with a model, part of LLM-as-judge.
Learn this properly