technical-term · Evaluation & verification · emerging · Reviewed
LLM as judge
The use of a language model to score, compare, classify, or critique another system’s output or trajectory.
Definition
An LLM judge applies a rubric to candidate outputs or agent traces and returns a score, preference, explanation, or failure label. Model graders can scale subjective evaluation, but they require calibration against experts and are not automatically independent, unbiased, or authorized to accept risk.
Use is widespread and evaluation methods are developing; bias, stability, contamination, and calibration remain material limitations.
Why it matters
A judge can measure qualities that deterministic tests cannot easily express.
The verification design must disclose judge model lineage, rubric, sampling, disagreement handling, and human calibration.
System anatomy
- Rubric
- Observable criteria and decision labels.
- Judge
- The model, version, context, and configuration performing evaluation.
- Calibration set
- Expert-scored examples used to measure agreement and bias.
- Disposition
- What the score is allowed to block, route, or inform.
Important distinctions
- Independent verifier
- A judge becomes one verifier method only when ownership, lineage, scope, trigger, block authority, and recorded disposition satisfy independence requirements.
- Deterministic test
- A model grader is probabilistic and rubric-dependent; a deterministic test has reproducible program logic.
Implementation signals
- Calibrate against domain experts
- Randomize or counterbalance comparison order
- Fuse model judgment with deterministic checks and human review
Failure modes
- Position or verbosity bias determines the winner
- Builder and judge share the same model lineage and blind spots
- A scalar score hides disqualifying failures
Sources and further study
- Anthropic — Demystifying evals for AI agents
A practical treatment of evaluating trajectories, outcomes, graders, tasks, and agent-environment interaction.
Use in this library: Official guidance. This is first-party engineering guidance drawn from Anthropic deployments. Evaluation designs remain task- and environment-specific, and the article does not establish independent verifier ownership by itself.
engineering · engineering · Published 2026-01-09 - arXiv — Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge
Documents position-related bias in model-based comparative judgment and motivates calibration and bias controls.
Use in this library: Empirical evidence. The study addresses selected judge models, prompts, and tasks; other grader failure modes and later systems require separate evaluation.
paper · paper · Published 2024-06-12 - NIST — Towards best practices for automated benchmark evaluations
Guidance on the integrity and limitations of automated benchmark evaluation practices.
Use in this library: Official guidance. Guidance describes recommended practice; citation does not prove that a control is implemented or effective in a particular environment.
guidance · guidance · Published 2026-01-30