/**
 * LLM-as-judge for the agent eval harness (R3.2).
 *
 * A single-call judge that scores a generated answer against a rubric — correct
 * objects, correct aggregation, honest trust label — returning a 0..1 score plus
 * a pass/fail. Per Anthropic's finding, a single call with a 0..1 score + pass/fail
 * is more consistent than elaborate multi-call rubrics. The `complete` callback is
 * injected so this is fully unit-testable offline and only reaches a real model
 * when the caller supplies a credentialed completion.
 */
export interface JudgeInput {
    question: string;
    sql?: string;
    answerText?: string;
    trustLabel?: string;
    resultSample?: unknown[];
    /** Optional gold reference (question intent / expected shape) for the rubric. */
    expectation?: string;
}
export interface JudgeVerdict {
    score: number;
    pass: boolean;
    rationale: string;
}
export type JudgeCompletion = (input: {
    system: string;
    user: string;
}) => Promise<string>;
export declare function judgeAnswer(input: JudgeInput, complete: JudgeCompletion): Promise<JudgeVerdict | undefined>;
export declare function parseJudgeVerdict(raw: string): JudgeVerdict | undefined;
/** Aggregate judge verdicts into a mean score and pass rate. */
export declare function summarizeJudgeVerdicts(verdicts: Array<JudgeVerdict | undefined>): {
    judged: number;
    meanScore: number | null;
    passRate: number | null;
};
//# sourceMappingURL=eval-judge.d.ts.map