/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   #511 base-consistency lint, GENERALIZED + COUNTRY-SCOPED (v2) — any synthetic slice vs the base.
 *
 *   Ported from scripts/lint-slice-vocab.py (pyarrow → @duckdb/node-api); behavior preserved
 *   byte-for-byte (same flags, same stdout, same verdicts). The base-root default routes through
 *   `dataRootPath` so the lab `/mnt/playpen` literal stays in its one home
 *   (core/utils/data-root.ts) and `$MAILWOMAN_DATA_ROOT` is honored; with the env unset it equals
 *   the Python default.
 *
 *   The #511 lesson: a synthetic slice must not label a token a tag the BASE dominantly labels
 *   something else, or training gets conflicting gradients on the same token and the minority (the
 *   slice) loses. This reads a slice's own (token -> tag) and checks each token against the base.
 *
 *   WHY v2 IS COUNTRY-SCOPED + FULL-COUNT (the night-2026-06-18 lesson, learned the hard way over
 *   three tries): a token's correct tag is COUNTRY-specific — "Paris" is locality in FR data and
 *   street in US "Paris Ave"; "Marion" is a US town AND many US "Marion" streets. So:
 *
 *   1. A cross-COUNTRY aggregate mis-judges any country-specific token (v1 uniform AND a proportional
 *        retry both false-flagged FR cities as "street" from US street-contexts).
 *   2. A SMALL sample is street-BIASED regardless, because the street sources (tiger 39 + nad 378 parts)
 *        dwarf the locality sources (a small US-scoped spot-check read Indianapolis 54% street vs
 *        its true 219700:29 LOCALITY). The fix: tally each slice token's base tag SCOPED to the
 *        country the slice uses it in (the base has a `country` column), over a LARGE/FULL scan
 *        (`fraction`, default 1.0). Pure-numeric tokens excluded (house_number/postcode are
 *        context-determined). An affix-split flag (slice street_suffix/_prefix vs base "street") is
 *        EXPECTED — the loader's affix-relabel handles it; weigh those separately.
 *
 *   Usage: mailwoman dev lint slice-vocab --slice <slice.parquet>
 *   [--base-version v0.5.0] [--base-root <dir>] [--fraction 1.0] [--threshold 0.7] [--min-count
 *   50]
 */
/**
 * Options for {@linkcode lintSliceVocab}.
 */
export interface LintSliceVocabOptions {
    /**
     * The slice parquet to lint.
     */
    slice: string;
    /**
     * Base corpus version. Default `v0.5.0`.
     */
    baseVersion?: string;
    /**
     * Base corpus root. Default `$MAILWOMAN_DATA_ROOT/corpus/versioned`.
     */
    baseRoot?: string;
    /**
     * Base-majority confidence floor for a contradiction. Default 0.7.
     */
    threshold?: number;
    /**
     * Minimum base support to judge a token. Default 50.
     */
    minCount?: number;
    /**
     * Fraction of base parts to scan (proportional per-source slice below 1.0). Default 1.0.
     */
    fraction?: number;
}
/**
 * One contradiction row: token, slice tag, base tag, base fraction, base total.
 */
export type SliceVocabRow = [token: string, sliceTag: string, baseTag: string, baseFrac: number, baseTotal: number];
/**
 * Findings summary returned by {@linkcode lintSliceVocab}.
 */
export interface LintSliceVocabSummary {
    /**
     * Real contradictions — the command exits 1 when nonzero.
     */
    errors: number;
    /**
     * Affix-split rows (EXPECTED — the loader's affix-relabel handles them).
     */
    warnings: number;
    findings: {
        contradictions: SliceVocabRow[];
        affixSplits: SliceVocabRow[];
    };
}
/**
 * Lint a synthetic slice's (token → tag) vocabulary against the base corpus, country-scoped.
 */
export declare function lintSliceVocab(options: LintSliceVocabOptions): Promise<LintSliceVocabSummary>;
//# sourceMappingURL=vocab.d.ts.map