/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   Inference-side gazetteer-anchor features (#464, knowledge-ladder rung 3.2) — the TS mirror of the
 *   Python training pipeline (`mailwoman_train/gazetteer_anchor.py`). Both consumers load the SAME
 *   codex-generated lexicon (`scripts/build-gazetteer-anchor-lexicon.mjs` →
 *   `data/gazetteer/anchor-lexicon-v1.json`) whose `rules` encode the match semantics as DATA, so
 *   the two implementations cannot drift. The model conditions on per-token candidate-tag-set clues
 *   fed alongside `input_ids`; this builds them from a raw address + its SentencePiece pieces.
 *
 *   The clue INFORMS, the model decides (model-first). `gazetteer-inference.test.ts` pins the matcher
 *   against the Python fixture: the homograph clue is symmetric, "in" ≠ "IN", multi-word countries
 *   paint every word.
 */
import type { TokenizedPiece } from "./tokenizer.ts";
/**
 * The candidate-tag-set feature width: country/region/po_box/cedex/homograph (the lexicon's slot count). Used for the
 * ONNX zero-fallback when a gazetteer-trained model is run with no clue data. MUST match the lexicon JSON's
 * `feature_dim` and the trained model's `gazetteer_feature_dim`.
 */
export declare const GAZETTEER_FEATURE_DIM = 5;
/**
 * Street-type evidence channel width (Option-A bundle) — the runner zero-fallback + lexicon feature_dim contract.
 */
export declare const STREET_TYPE_FEATURE_DIM = 1;
/**
 * Locality-surface evidence channel width (locality + locality_homograph).
 */
export declare const LOCALITY_SURFACE_FEATURE_DIM = 2;
/**
 * The loaded lexicon — the JSON shape from build-gazetteer-anchor-lexicon.mjs.
 */
export interface GazetteerLexicon {
    featureDim: number;
    slots: readonly string[];
    bits: Record<string, number>;
    maxNgram: number;
    /**
     * Case-insensitive: key = word_norm lowercased → bitmask.
     */
    entries: Map<string, number>;
    /**
     * Case-SENSITIVE: key = word_norm uppercased → bitmask (surface must already be uppercase).
     */
    codeEntries: Map<string, number>;
    /**
     * V3.23 digit guard (`rules.digit_guard`): a matched span paints NOTHING when any span word or the nearest non-empty
     * neighbor word carries a decimal digit — evidence painted beside a house number swallowed the digit into the span.
     * Rides the lexicon so train/inference stay symmetric by construction; false on pre-v3.23 artifacts.
     */
    digitGuard: boolean;
}
/**
 * Parse the lexicon JSON (already `JSON.parse`d — keeps this module browser-safe; caller reads).
 */
export declare function parseGazetteerLexicon(raw: {
    feature_dim: number;
    slots: string[];
    bits: Record<string, number>;
    max_ngram: number;
    entries: Record<string, number>;
    code_entries: Record<string, number>;
    rules?: {
        digit_guard?: boolean;
    };
}): GazetteerLexicon;
/**
 * Scan the raw surface and paint each char with its candidate-tag bitmask (mirrors Python).
 */
export declare function gazetteerCharPaint(text: string, lexicon: GazetteerLexicon): number[];
/**
 * Channel choreography (#464, v0.9.13 postcode fix; DeepSeek 2026-06-10): zero the gazetteer clue on pieces within
 * `window` of a postcode-anchor hit. The clue fires on the region token (`CA`/`GA`) immediately before a US postcode;
 * its additive vector strengthens `B-region`, which makes the `B-region → B-postcode` CRF transition less competitive
 * and drops the postcode (~3pp, US-only — FR postcode precedes the locality, no region neighbor). Suppressing the clue
 * adjacent to the postcode removes the interference while leaving every other clue intact. Returns a NEW
 * features/confidence pair (does not mutate). `anchorConfidence[i] > 0` marks postcode-span pieces. PAIRS WITH the
 * train-time half (`gazetteer_anchor.suppress_gazetteer_near_postcode`) — enable both or neither.
 */
export declare function suppressGazetteerNearPostcode(gazetteer: {
    features: number[][];
    confidence: number[];
}, anchorConfidence: ReadonlyArray<number>, window?: number): {
    features: number[][];
    confidence: number[];
};
/**
 * Per-piece gazetteer features + confidence for `text`, projected onto its SP `pieces` by the SAME char→piece rule the
 * labels use (a piece takes the bits of the first non-whitespace char it covers). Returns `(pieces × featureDim)`
 * features + `(pieces,)` confidence (1.0 wherever any bit fires).
 */
export declare function buildGazetteerFeatures(text: string, pieces: ReadonlyArray<TokenizedPiece>, lexicon: GazetteerLexicon): {
    features: number[][];
    confidence: number[];
};
//# sourceMappingURL=gazetteer-inference.d.ts.map