/**
 * How much of a score is signal.
 *
 * A suite reported a mean and nothing else, so two experiments differing by three
 * points read as a difference. At the n a hand-built suite has, that is
 * usually noise, and there was no number on the page that would have said
 * so.
 *
 * Evan Miller, "Adding Error Bars to Evals" (arXiv:2411.00640), is the
 * reference. Two of its results shape what is and is not computed here.
 */
export interface ScoreUncertainty {
    /** Cases that produced a score. Not the number of cases run. */
    readonly n: number;
    /** Sample standard deviation, Bessel-corrected. */
    readonly stdDev: number;
    /** Standard error of the mean. */
    readonly stdError: number;
    /** Half-width of the 95% interval: the mean plus or minus this. */
    readonly margin95: number;
    /**
     * The 95% interval, clamped to the score range.
     *
     * Clamped because a mean of 0.95 with a wide interval otherwise reports
     * an upper bound above 1, which is not a possible score and makes a
     * reader distrust the whole figure. The clamp is cosmetic and the
     * margin above is not — read that one for the width.
     */
    readonly ci95: readonly [low: number, high: number];
    /**
     * True when there is not enough data for an interval at all.
     *
     * One case has no spread to measure. Reporting `±0` there would be the
     * most confident-looking output the suite can produce, from the least
     * evidence it can have.
     */
    readonly undefinedInterval: boolean;
}
/**
 * Spread of a set of scores, with the interval a reader should apply.
 *
 * **Assumes the cases are independent, and they may not be.** Miller's
 * clustered standard errors run up to 3x the naive figure when cases come
 * in related groups — several cases derived from one scenario, or one
 * document, or one seed. This harness has no grouping key on a case, so
 * there is nothing here to cluster on and this returns the naive figure.
 * Where a suite does build several cases from one source, treat the
 * interval below as a floor rather than as the answer.
 *
 * Stated rather than silently assumed because a too-narrow interval is
 * worse than none: it turns "we cannot tell" into a number that looks
 * like we can.
 */
export declare function uncertaintyOf(scores: readonly number[]): ScoreUncertainty;
/**
 * One line a reader can act on, for a surface that prints a score.
 *
 * Names the interval rather than only the mean, because the mean alone is
 * the thing that has been over-read. An interval spanning most of the
 * scale says the suite cannot currently tell two experiments apart, and that is
 * the most useful sentence such a suite can produce.
 */
export declare function describeUncertainty(mean: number, u: ScoreUncertainty): string;
//# sourceMappingURL=uncertainty.d.ts.map