import type { LLMProvider } from '../types/provider/interface.js';
import type { Scorer } from './types.js';
/**
 * Grade an open-ended answer with a model.
 *
 * Every other scorer here is a pure function over the turn, which is what
 * makes them cheap and reproducible — and also what makes them unable to
 * say anything about whether an answer is *good*. `containsScorer` can
 * check that a required phrase appears; it cannot tell a correct
 * explanation from a fluent wrong one. That was the hole: the dimension
 * most worth guarding was the one with no scorer behind it.
 *
 * A judge is a network call, so this scorer can fail in ways a pure
 * function cannot. It throws rather than returning a low score, and the
 * harness files a throw as `unavailable` — a rate limit must not read as
 * a regression. See `Score.unavailable`.
 */
export interface JudgeScorerConfig {
    /** Defaults to `'judge'`. Give it a distinct name to run several. */
    name?: string;
    provider: LLMProvider;
    model: string;
    /**
     * What "good" means, in the caller's words. REQUIRED.
     *
     * A judge asked to rate quality with no rubric rates fluency, which
     * correlates with almost nothing worth measuring and drifts whenever
     * the judge model changes. Making this optional would make the scorer
     * easy to use and its output meaningless, so it is not optional.
     */
    rubric: string;
    /**
     * Highest grade on the scale. Default 4.
     *
     * An integer scale, not a 0..1 float: models place a continuous score
     * poorly and cluster on round numbers, while a short ordinal scale
     * against a written rubric is a judgement they can actually make. The
     * result is divided down to 0..1 for the report.
     *
     * The default is EVEN on purpose — an odd scale has a midpoint, and a
     * midpoint is where an uncertain judge parks. Forcing a side produces a
     * signal; a pile of 3-out-of-5s does not.
     */
    scale?: number;
    /**
     * Show the judge which tools the turn called. Default false.
     *
     * Useful when the rubric is about method rather than answer, and a
     * needless cost otherwise — the trajectory is usually longer than the
     * answer it produced.
     */
    includeTrajectory?: boolean;
    /**
     * Cap on the answer text handed to the judge. Default 20000.
     *
     * Truncation is disclosed IN the prompt. A judge shown a silently cut
     * answer marks it down for stopping mid-sentence, which scores our
     * truncation rather than the turn.
     */
    maxOutputChars?: number;
    /**
     * Milliseconds without a judge chunk before its provider stream is
     * treated as stalled. Defaults to the SDK's finite provider-stream idle
     * bound. Set `0` only for explicit unbounded compatibility.
     */
    streamIdleTimeoutMs?: number;
}
export declare function judgeScorer(config: JudgeScorerConfig): Scorer;
//# sourceMappingURL=judge.d.ts.map