/**
 * Classification metrics for the System One benchmark (T12495).
 *
 * Two-label sites report precision, recall, F1 and false-positive rate for
 * their positive class. The multi-class observation-type site reports macro
 * averages over the labels present in the gold set (one-vs-rest per label).
 *
 * @task T12495
 * @epic T12486
 */
/** One scored prediction. */
export interface BenchPrediction {
    /** Gold label. */
    readonly gold: string;
    /** Predicted label. */
    readonly predicted: string;
}
/** Classification quality over a set of predictions. */
export interface BenchClassificationMetrics {
    /** Scored predictions. */
    readonly n: number;
    /** Share predicted correctly. */
    readonly accuracy: number | null;
    /** Precision (positive class, or macro average). */
    readonly precision: number | null;
    /** Recall (positive class, or macro average). */
    readonly recall: number | null;
    /** F1 (positive class, or macro average). */
    readonly f1: number | null;
    /** False-positive rate (positive class, or macro average one-vs-rest). */
    readonly falsePositiveRate: number | null;
}
/**
 * Score predictions.
 *
 * @param preds - Gold and predicted labels.
 * @param positive - The positive class of a two-label site; `null` → macro averages.
 * @returns The metrics (`null` where undefined, e.g. no positives predicted).
 */
export declare function classificationMetrics(preds: readonly BenchPrediction[], positive: string | null): BenchClassificationMetrics;
/** p50 and p95 of latencies, in ms. */
export interface BenchLatency {
    /** Median. */
    readonly p50Ms: number | null;
    /** 95th percentile. */
    readonly p95Ms: number | null;
}
/**
 * Latency percentiles.
 *
 * @param samples - Latencies in ms.
 * @returns p50 and p95 (nearest rank).
 */
export declare function latencyOf(samples: readonly number[]): BenchLatency;
//# sourceMappingURL=metrics.d.ts.map