The written standard a grader applies, human or machine. Vague criteria produce inconsistent scores; specific criteria with worked examples produce useful ones. Writing a good rubric forces you to say what you actually want, which is often the first time anyone on the project has done so.
Technical view
Effective rubrics decompose quality into independently assessable dimensions such as factual accuracy, instruction compliance, completeness and tone, with defined levels and anchor examples for each. Binary checklists usually outperform Likert scales for reliability. When used with a model judge, the rubric doubles as the prompt, and its wording measurably affects scores, so it should be versioned and validated against human ratings like any other instrument.
Related
Part of AI Terminology Index, section: Measuring Performance.