C.W.K.
Stream
Lesson 04 of 10 · published

LLM-as-Judge — Use Cases and Pitfalls

~16 min · evaluation, judges

Level 0Apprentice
0 XP0/100 lessons0/14 achievements
0/120 XP to next level120 XP to go0% complete

Cheap, fast, biased

Using an LLM to grade other LLM outputs is the cheapest way to scale evaluation. It's also the most easily fooled. Used carefully, it accelerates work; used carelessly, it produces a comforting number that doesn't predict real quality.

Where LLM judges work

  • Format checking — is this JSON valid? does the response include all required fields?
  • Pairwise comparison with a clear rubric.
  • Categorical classification (refund vs complaint vs feature request).
  • Hallucination detection against a reference ("does the response contradict the source?").

Where LLM judges fail

  • Numerical scoring on subjective qualities (1–5 helpfulness scores are noisy).
  • Anything where the judge has the same blind spots as the model under test.
  • Adversarial situations — a judge model can be prompt-injected too.

Calibrate the judge

Spot-check 20% of judge decisions with a human. If agreement rate is below 85%, the judge isn't usable for that task; tighten the rubric or do more human eval.

Code

Hallucination judge·markdown
## Task
Given the source documents and the response, decide whether the response contains any claim not supported by the source.

## Output
{
  "hallucinated": true | false,
  "unsupported_claims": ["<verbatim claim text>", ...],
  "verdict_confidence": "high" | "medium" | "low"
}

## Source
{{source}}

## Response
{{response}}

External links

Exercise

Build an LLM judge for a specific task in your system. Calibrate against 30 human-graded outputs. If agreement is below 85%, tighten the rubric or replace with categorical scoring.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.