/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   Pre-compute corpus-wide token + bigram label distributions for the corpus linter.
 *
 *   Reads one or more Parquet shards, builds per-(token, label) and per-(bigram, label-bigram)
 *   histograms, and serializes them as JSON. The output file is consumed by `lint-corpus-shard.ts`
 *   as the baseline against which a new shard is compared.
 *
 *   Stats are cheap to compute (~5–30s per 100K rows) but expensive enough that we cache them between
 *   linter invocations. Re-run this script whenever the corpus changes substantially (a new
 *   mainline shard added, a source-pool re-weighted, etc.).
 *
 *   Output schema:
 *
 *   ```ts
 *   interface CorpusStats {
 *     row_count: number
 *     shard_paths: string[]
 *     tokens: { [token: string]: { [label: string]: number } }
 *     bigrams: { [token_bigram: string]: { [label_bigram: string]: number } }
 *     // token_bigram = "tok1tok2" (US sep), label_bigram = "lab1lab2"
 *     // For memory: only keep bigrams with count >= MIN_BIGRAM_COUNT (2).
 *   }
 * ```
 *
 *   Usage: node scripts/build-corpus-stats.ts\
 *   --shards <glob-pattern-or-dir>\
 *   --output <stats.json>
 *
 *   For a quick local-corpus baseline (limited but useful for linter testing): node
 *   scripts/build-corpus-stats.ts\
 *   --shards $MAILWOMAN_DATA_ROOT/corpus/versioned/v0.4.0/corpus-v0.4.0/train/\
 *   --output /tmp/corpus-stats-local.json
 */
export interface CorpusStatsOptions {
    shardsArg: string;
    outputPath: string;
    limitPerShard?: number;
}
export declare function buildCorpusStats(args: CorpusStatsOptions): Promise<void>;
//# sourceMappingURL=corpus-stats.d.ts.map