/**
 * panel-grader — multi-judge LLM grader with consensus aggregation.
 *
 * Runs the same rubric prompt against N judge models in parallel, then
 * combines their normalized scores with a configurable aggregation
 * strategy (majority / unanimous / mean / median / min).
 *
 * Re-uses the same tool schema, system message, and per-criterion
 * structure as `prompt-grader` (via `buildRubricJudge` and `buildUserMessage`)
 * so reporters and downstream tooling see one consistent "judge response"
 * shape per judge.
 *
 * Design notes:
 * - Judge-error strictness: any judge that fails (non-zero exit / crash /
 *   timeout / malformed response) forces the whole panel to fail, regardless
 *   of the surviving judges' consensus. The surviving judges still produce the
 *   reported `score`, but `passed` is forced `false` and the failed judges are
 *   surfaced in `metadata.failed_judges`.
 * - Concurrency: all judges fan out together via `Promise.allSettled`. The
 *   panel grader intentionally does not expose a per-invocation concurrency
 *   knob — LLM rate-limit / throughput control belongs at a global layer
 *   (e.g. on the `LlmClient` or runner) so it applies uniformly across every
 *   LLM-using grader rather than only within a single panel call.
 * - Disagreement: the spread (`max − min`) of judge scores is exposed in
 *   `metadata.disagreement` as a raw primitive. The grader intentionally
 *   does not flag it against a threshold or affect `passed`/`score` —
 *   downstream tooling (reporters, dashboards, alerts) can apply its own
 *   policy on top of the raw value.
 */
import type { Grader, GraderInput, GraderMetadata, GraderResult } from "../types.js";
import type { LlmClient } from "./types.js";
export declare class PanelGrader implements Grader {
    metadata: GraderMetadata;
    private client;
    constructor(client: LlmClient);
    /**
     * `panel "did it use bicep"`, else `panel gpt-5,claude@high` — the rubric is
     * the clearest discriminator; the judge lineup is the fallback since `models`
     * is the only required config.
     *
     * A judge's `reasoning_effort` is part of that lineup: two panels differing
     * only by effort run genuinely different judges, so omitting it would derive
     * one name for both and drop them back to positional `#n` labels. Uses the
     * same `model@effort` spelling as {@link judgeLabels}.
     */
    defaultName(config: Record<string, unknown>): string;
    grade(input: GraderInput): Promise<GraderResult>;
}
//# sourceMappingURL=panel-grader.d.ts.map