/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   `mailwoman corpus audit` — per-source shard-count vs source_weight diagnostic.
 *
 *   Reads a corpus dir's MANIFEST.json (or scans shards directly), counts shards per source,
 *   optionally loads a training config to pair the counts with the configured source_weights, and
 *   reports the estimated sampled-row distribution at training time.
 *
 *   Would have caught v0.3.0's "NAD = 411/674 train shards × 2.0 weight = ~75% of sampled mix"
 *   finding before the v0.3.0 retrospective surfaced it.
 *
 *   Emits warnings to stderr and the audit table to stdout; never throws on an empty corpus.
 */
export interface AuditOpts {
    corpusDir: string;
    configPath?: string;
    /**
     * Sample at most N shards per split when counting sources. Default 100 for speed; bump to read the full set on a slow
     * run. The first row of each shard determines its source — corpus-v0.2.0+ shards are 100% source-segregated, so a
     * one-row read is authoritative.
     */
    sampleShardCount?: number;
}
export declare function audit(opts: AuditOpts): void;
//# sourceMappingURL=audit.d.ts.map