/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   Evaluate the #244 coarse-placer: in-distribution accuracy + per-class + calibration (ECE) on the
 *   held-out test split, AND the abstention story on the multi-script set — off-map scripts
 *   (Cyrillic, Arabic, Thai, …, none of them in the 11 trained countries) SHOULD draw low
 *   confidence → abstain, which is the "probably off my loaded map" behavior the design wants.
 *
 *   Run: `mailwoman placer eval in-distribution [--model <dir>] [--abstain 0.5]`
 */
/**
 * Options for {@linkcode evalCoarsePlacer}.
 */
export interface EvalCoarsePlacerOptions {
    /**
     * Model artifact dir. Default `$MAILWOMAN_DATA_ROOT/coarse-placer/model`.
     */
    model?: string;
    /**
     * Abstention threshold. Default 0.5.
     */
    abstain?: number;
    /**
     * Dataset dir (`test.jsonl`). Default `<repo>/data/coarse-placer`.
     */
    data?: string;
}
/**
 * Result of {@linkcode evalCoarsePlacer}.
 */
export interface EvalCoarsePlacerResult {
    n: number;
    /**
     * Overall accuracy in percent.
     */
    accuracy: number;
    /**
     * 10-bucket expected calibration error.
     */
    ece: number;
}
/**
 * Coarse-placer in-distribution eval — see the module doc. Emits the report to stdout.
 */
export declare function evalCoarsePlacer(options?: EvalCoarsePlacerOptions): Promise<EvalCoarsePlacerResult>;
//# sourceMappingURL=eval.d.ts.map