Compare Architectures

Side-by-side: use case, requirements, latency, cost, strengths, weaknesses, failure modes, example — and when to choose each.

LLM-as-judgeDeterministic evaluators
Use caseOpen-ended quality: helpfulness, faithfulness, toneVerifiable properties: schema, exact match, tool called, length
RequirementsRubric, judge prompt, calibration set graded by humansCode assertions and reference values
LatencyOne extra model call per sampleMilliseconds
CostTokens per evaluation; expensive at scaleNear zero
StrengthsScales human-like judgement to thousands of samplesReproducible, cheap, no drift
WeaknessesBiased (position, verbosity, self-preference); noisyCannot assess quality beyond what is checkable
Failure modesJudge agrees with confident wrong answersPassing tests while answers are useless
ExampleIs this summary faithful to the source?Did the output parse as valid JSON with required fields?
Choose this whenChoose an LLM judge for qualities you cannot express as code, and calibrate it against human labels first.Choose deterministic evaluators for everything checkable; run them first, on every change.