/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   Soft-prior emission biases derived from `QueryShape`.
 *
 *   When the QueryShape sub-system has identified a known-format span (US ZIP, UK postcode, PO box,
 *   etc.), this module produces an additive bias matrix that nudges the encoder's per-token
 *   emissions toward the matching BIO label. The biases compose with the structural BIO mask in the
 *   Viterbi decoder — confident encoder predictions still win, but uncertain ones get pulled toward
 *   the format-implied label.
 *
 *   Bitter-lesson-safe boundary: we don't override the encoder, just bias it. The encoder remains the
 *   authority on context-dependent calls (the "Buffalo Wild Wings, Buffalo, NY" disambiguation);
 *   the QueryShape prior helps on the easy cases (a 5-digit token is _probably_ a postcode).
 *
 *   RETIRED 2026-07-17 — the LOCALITY bias (regionAbbreviations → boost B/I-locality on preceding
 *   tokens). The M1 stack ablation (docs/articles/evals/2026-07-17-m1-stack-ablation.md) measured the
 *   full prior at −2.3 micro / −7.8 locality on golden-us, and the three-arm sub-ablation attributed
 *   100% of the damage to the locality half: stripping it recovered locality exact-match 0.7822 →
 *   0.8546 (= the no-prior arm), while the known-format half was exactly neutral. The failure mode was
 *   venue/org absorption on registry-style rows ("DANVILLE HEALTH CENTER, 26 Cedar Lane, Danville VT"
 *   → locality "danville health center"): the backward walk from a detected region abbreviation
 *   crossed comma gaps and dragged venue text into locality. The WOF bare-name over-emission it was
 *   built to counter no longer reproduces — the model outgrew it (same lifecycle as the #956-era
 *   near-postcode suppression, also measured negative in M1). The known-format boosts below remain.
 *
 *   Uses structural typing for the QueryShape input so this module has zero dependencies on
 *   `@mailwoman/query-shape` — consumers compute the shape with that package, pass it in here.
 */
export interface QueryShapeLike {
    knownFormats: ReadonlyArray<KnownFormatHitLike>;
    regionAbbreviations?: ReadonlyArray<RegionAbbreviationHitLike>;
}
interface RegionAbbreviationHitLike {
    start: number;
    span: string;
}
export interface KnownFormatHitLike {
    format: string;
    span: {
        start: number;
        end: number;
    };
    /**
     * 0..1; ambiguous patterns (e.g. 5-digit US/FR/DE overlap) score lower.
     */
    confidence: number;
}
/**
 * Minimal subset of `TokenizedPiece` this module consumes.
 */
export interface TokenLike {
    start: number;
    end: number;
}
export interface BuildPriorsOpts {
    /**
     * Maximum bias magnitude (in log-odds units). Default 1.0 — adds up to ~e^1 ≈ 2.7× odds to the favored label.
     * Confidence-scaled, so a 0.6-confidence format hit gets +0.6 max bias.
     */
    biasScale?: number;
    /**
     * Raw input text — enables the SCOPED locality bias (bare admin doubletons only; see `applyScopedLocalityBias`).
     * Without it the digit guard cannot run, so the locality bias never fires.
     */
    inputText?: string;
}
/**
 * Build a `[seqLen][numLabels]` matrix of additive log-bias to be added to encoder emissions before Viterbi decoding.
 *
 * For each (token, format-hit) pair where the token's character span overlaps the hit's span, the matrix entry for the
 * format's mapped label receives `hit.confidence × biasScale`. Tokens that don't overlap any hit, or for which no label
 * mapping exists, get 0.
 *
 * Returns the all-zeros matrix if `shape.knownFormats` is empty — composes harmlessly.
 */
export declare function buildEmissionPriors(shape: QueryShapeLike, tokens: ReadonlyArray<TokenLike>, labels: ReadonlyArray<string>, opts?: BuildPriorsOpts): number[][];
/**
 * Element-wise add two matrices of equal shape. Returns a new matrix.
 */
export declare function addEmissionMatrix(emissions: number[][], priors: number[][]): number[][];
export {};
//# sourceMappingURL=query-shape-prior.d.ts.map