/**
 * Mixed grading: deterministic assertions + LLM judge + multi-dimensional scoring.
 */
import type { GradeResult, ExecutorFn, JudgeConfig, Sample, ToolCallInfo, TurnInfo } from '../types/index.js';
interface GradeOptions {
    output: string;
    sample: Sample;
    /**
     * Unified judge config — always non-empty. `length === 1` runs the single-judge
     * quick path (one judge per (sample × dimension)), `length >= 2` runs the
     * ensemble path with per-judge breakdown + Pearson/MAD agreement metrics
     * (refutes "judge same-modality bias" critique).
     *
     * The first entry (`judgeModels[0]`) is also used as the "primary judge" for
     * single-call paths (async assertions): one LLM call per assertion, no ensemble
     * aggregation — the primary judge is a deterministic representative.
     */
    judgeModels: JudgeConfig[];
    /**
     * Map `executor name → ExecutorFn`. The map only needs entries for executors
     * actually invoked by this grade call — i.e. when the sample triggers an LLM
     * code path (async assertions, dimension scoring, single rubric, ensemble
     * member). Sync-only samples can pass `{}`. Failures are lazy: missing entry
     * throws only at the call site, not on grade entry.
     */
    judgeExecutors?: Record<string, ExecutorFn>;
    allowLlmJudge?: boolean;
    execMetrics?: {
        costUSD?: number;
        durationMs?: number;
        numTurns?: number;
        toolCalls?: ToolCallInfo[];
        turns?: TurnInfo[];
        mockStats?: {
            hits: number;
            misses: number;
            perMock: Record<string, number>;
        };
    };
    samplesDir?: string;
    /**
     * Run the LLM judge N times per (sample × dimension) pair and report mean + stddev.
     * Default 1 (single judge call). Useful for measuring judge self-consistency:
     * a high stddev means the judge isn't stable on this rubric and the score is noisy.
     */
    judgeRepeat?: number;
    /**
     * length-debias toggle. Defaults to true — judge prompt includes the
     * "length is not a quality signal" instruction. Set false (via `--no-debias-length`)
     * to drop it (the debias-off prompt variant) for reproducing older no-length-debias
     * reports. The presentation/tone neutrality instruction is always on regardless.
     */
    lengthDebias?: boolean;
}
/**
 * Grade a model output against a sample's criteria.
 */
export declare function grade({ output, sample, judgeModels, judgeExecutors, allowLlmJudge, execMetrics, samplesDir, judgeRepeat, lengthDebias }: GradeOptions): Promise<GradeResult>;
export { runAssertions, validateJsonSchema } from './assertions.js';
export { buildTraceSummary } from './judge.js';