import type { RunSummary } from "../runDirectory/list.js";
/**
 * Statistics over one batch: one suite invocation where every test ran the
 * same number of trials. A test's mean and standard error come from its
 * trials; the batch's accuracy is the mean over every scored run, and its
 * standard error is taken over the per-trial batch means, paired by trial
 * index (trial 1 of every test is one sample), the way a leaderboard reports
 * "accuracy ± SE over k trials". Nothing here is stored: it is recomputed
 * from run summaries wherever they live.
 */
export type TestStatistics = {
    testId: string;
    trials: number;
    /** Mean score over the scored trials; null when none has a score. */
    mean: number | null;
    /** Sample standard deviation / √n; null below two scored trials. */
    standardError: number | null;
    meanCostUsd: number;
    meanDurationMs: number;
};
export type BatchStatistics = {
    /** The shared batch id; null for a lone run from before batches existed. */
    batch: string | null;
    tests: TestStatistics[];
    trials: number;
    /** Mean score over every scored run in the batch; null when none has one. */
    accuracy: number | null;
    /** Sample standard deviation of the per-trial batch means / √k; null below two trials. */
    standardError: number | null;
    totalCostUsd: number;
    totalDurationMs: number;
};
/** Split runs into batches by their batch id (a run without one is a batch
 *  of its own), in first-seen order, and compute each batch's statistics. */
export declare function batchStatisticsByBatch(runs: readonly RunSummary[]): BatchStatistics[];
/** A batch that could not be scored as a whole (an uneven trial grid — a run
 *  cut short by the batch cost cap or an interrupt — or mixed ids), with the
 *  reason `batchStatistics` gave. */
export type IncompleteBatch = {
    batch: string | null;
    reason: string;
};
/** Like `batchStatisticsByBatch`, but a batch that `batchStatistics` refuses
 *  is reported rather than thrown, so one incomplete batch never hides the
 *  others' statistics. */
export declare function batchStatisticsByBatchTolerant(runs: readonly RunSummary[]): {
    batches: BatchStatistics[];
    incomplete: IncompleteBatch[];
};
/** Statistics for exactly one complete batch. Refuses an empty input, runs
 *  from different batches, more than one run without a batch id, and any
 *  incomplete trial grid: every test must have trials exactly `1..k` for one
 *  `k`, so a pairing by trial index is always between real trials. */
export declare function batchStatistics(runs: readonly RunSummary[]): BatchStatistics;