Your LLM Judge Needs a Regression Test Suite
rubric-bench v0.1 is an open-source tool that regression-tests LLM judges against golden sets, catching grading drift in CI before it reaches production.
This post argues that LLM-based judges — graders, triage classifiers, eval scoring models — are routinely shipped to production without tests, even though their behavior can silently shift with model updates, provider changes, or prompt edits. The proposed fix is rubric-bench v0.1, a TypeScript library and CLI that treats a grading model like any other tested code: it runs a grader against a versioned golden set of cases, reports accuracy, diffs two runs to catch regressions, and can gate CI builds when a previously passing case starts failing.
In the first real run against a 72-case introductory-statistics golden set, the grader scored 95.8%, with all three misses falling into the same gray zone where human graders also disagree — partial-credit answers marked as incorrect. The author also shares practical lessons: capturing per-case API errors instead of failing the whole run, and treating LLM-judge non-determinism (a case or two flipping between identical runs) as noise unless a pattern emerges.
Guidance for building your own golden set includes using discrete verdicts instead of continuous scores, sourcing real inputs before writing synthetic ones, keeping partial-credit cases genuinely ambiguous, tagging every case, and freezing golden-set versions once cited in a report. The author is clear that rubric-bench is regression testing, not psychometric validation — it shows that a judge changed, not that it agrees with human raters.
The engineering takeaway is straightforward: if an LLM's judgment gates something users care about, it's a production dependency that deserves the same discipline as schema migrations — pin the version, benchmark before swapping, and gate prompt changes through CI.