{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://raw.githubusercontent.com/api-evangelist/evals/refs/heads/main/json-schema/evals-judge-schema.json",
  "title": "Judge",
  "description": "A specialized scorer that uses an LLM (or human panel) to render judgment on the system-under-test's output. Captures the judge model, the judge prompt, the rubric, and calibration data — the things you'd want when comparing one judge to another or validating a judge against human ratings. Patronus AI's Lynx and GLIDER, OpenAI's model_graded evals, and TruLens feedback functions are all instances of this shape.",
  "type": "object",
  "properties": {
    "id": {
      "type": "string",
      "example": "judge_groundedness_panel_v1"
    },
    "name": {
      "type": "string",
      "example": "Groundedness Judge"
    },
    "description": {
      "type": "string",
      "example": "LLM-as-a-judge scoring whether an answer is grounded in retrieved context."
    },
    "judge_kind": {
      "type": "string",
      "enum": ["llm", "human", "model_panel", "human_panel", "distilled_evaluator"],
      "example": "llm"
    },
    "model": {
      "type": "object",
      "description": "Model used as the judge (when judge_kind is llm or distilled_evaluator).",
      "properties": {
        "provider": { "type": "string", "example": "anthropic" },
        "name": { "type": "string", "example": "claude-opus-4-7" },
        "version": { "type": "string" }
      },
      "required": ["provider", "name"]
    },
    "prompt_template": {
      "type": "string",
      "description": "The prompt used by the judge. Should include slots for {input}, {output}, {expected}, {context} as applicable.",
      "example": "You are a strict evaluator. Given the question {input}, the retrieved context {context}, and the answer {output}, score groundedness on a 0-1 scale. Explain your reasoning."
    },
    "rubric": {
      "type": "string",
      "description": "Human-readable rubric the judge follows."
    },
    "output_format": {
      "type": "string",
      "enum": ["score", "score_and_rationale", "label", "label_and_rationale", "pairwise_preference"],
      "example": "score_and_rationale"
    },
    "calibration": {
      "type": "object",
      "description": "Optional calibration evidence — agreement with human raters, Cohen's kappa, etc.",
      "properties": {
        "human_agreement": { "type": "number", "example": 0.87 },
        "kappa": { "type": "number", "example": 0.74 },
        "sample_size": { "type": "integer", "example": 250 },
        "calibrated_on": { "type": "string", "format": "date-time" }
      }
    },
    "tags": {
      "type": "array",
      "items": { "type": "string" },
      "example": ["rag", "groundedness", "llm-judge"]
    }
  },
  "required": ["id", "name", "judge_kind"]
}