/**
 * SemanticNoveltyEncoder — the learned-local-semantic ADVISORY layer for the
 * conjecture novelty check (D.060 / scope 2026-05-23_holoembed-semantic-encoder).
 *
 * The shipped novelty check (`assessConjectureNovelty`) uses HoloEmbed char-trigram
 * histograms — a LEXICAL exact-restatement guard that catches near-verbatim restatements
 * only (a paraphrase of a known result returns `novel`, a false negative — the W.520 trap).
 * This module adds a genuinely SEMANTIC layer using a local learned model
 * (`Xenova/all-MiniLM-L6-v2`) via the already-present `@huggingface/transformers` — no new
 * heavy dependency, runs offline/CPU (sovereign, F.063), NOT an external API.
 *
 * LAYERING (the determinism resolution): the deterministic trigram guard stays
 * RECEIPT-BINDING; this learned layer is ADVISORY only — it does NOT feed any receipt
 * hash. Same-text determinism is verified (byte-identical vectors across runs on one
 * machine), but cross-fleet floating-point reproducibility is unproven, so it must not
 * enter a receipt until that determinism gate passes. It returns a review signal
 * ("resembles known result X at 0.82"), never a receipt-binding verdict.
 *
 * Empirically (probe 2026-05-23, this model, cached): Euler's-formula claim vs a paraphrase
 * = cos 0.73; vs an unrelated sentence = cos 0.08 — clean separation the trigram check (at
 * 0.995) misses entirely. Async (model inference); the trigram path stays sync.
 */
import type { ConjecturePriorArtEntry } from './ConjectureEngine';
/** Pinned model — local ONNX via transformers.js. 384-dim, mean-pooled, L2-normalized. */
export declare const SEMANTIC_NOVELTY_MODEL: "Xenova/all-MiniLM-L6-v2";
/**
 * Default advisory threshold, CALIBRATED (scope P2, 2026-05-23) on the labeled eval set
 * below: same-result paraphrases scored 0.65–0.72, different/unrelated scored 0.02–0.32
 * (cleanly separable, gap 0.34, midpoint ≈0.49). 0.5 sits centered in the gap — 0.18
 * above every no-match and 0.15 below every match — so it is robust to harder paraphrases
 * and recall-favoring (right for an ADVISORY flag, where a missed rediscovery is worse
 * than a flagged-for-review false positive). Still advisory-only; re-run calibration if
 * the model changes.
 */
export declare const SEMANTIC_NOVELTY_THRESHOLD = 0.5;
export type SemanticNoveltyStatus = 'near-duplicate' | 'novel';
export interface SemanticNoveltyMatch {
    priorArtId: string;
    source: string;
    title?: string;
    statement: string;
    similarity: number;
}
export interface SemanticNoveltyAssessment {
    /** ADVISORY — never receipt-binding (see module note). */
    binding: 'advisory';
    provider: 'transformers.js';
    modelId: typeof SEMANTIC_NOVELTY_MODEL;
    status: SemanticNoveltyStatus;
    threshold: number;
    corpusSize: number;
    query: string;
    nearest: SemanticNoveltyMatch | null;
}
/** Embed text into a mean-pooled, L2-normalized vector via the local learned model. */
export declare function embedSemantic(text: string): Promise<number[]>;
export interface LabeledNoveltyPair {
    query: string;
    reference: string;
    /** true = query is the SAME result as reference (a paraphrase); false = different/unrelated. */
    isMatch: boolean;
}
/**
 * Labeled eval set for threshold calibration (scope P2) — paraphrases of known results
 * across the engine's live domains (should match) vs different/unrelated claims (should not).
 * Durable artifact: reused by the determinism gate (P1) and corpus work (P3).
 */
export declare const LABELED_NOVELTY_EVAL_SET: ReadonlyArray<LabeledNoveltyPair>;
export interface ThresholdCalibration {
    threshold: number;
    separable: boolean;
    minMatchSim: number;
    maxNoMatchSim: number;
    gap: number;
}
/**
 * Pure calibration: given scored (similarity, isMatch) pairs, return the midpoint
 * threshold between the lowest match and the highest no-match, plus whether the classes
 * are separable and the margin. Deterministic; no model required.
 */
export declare function calibrateNoveltyThreshold(scored: ReadonlyArray<{
    similarity: number;
    isMatch: boolean;
}>): ThresholdCalibration;
/** Cosine similarity of two equal-length vectors (inputs are L2-normalized → dot product). */
export declare function cosineSimilarity(a: ReadonlyArray<number>, b: ReadonlyArray<number>): number;
/**
 * ADVISORY semantic novelty assessment: embeds the query + each corpus statement with the
 * local learned model and flags `near-duplicate` if the nearest match is >= threshold.
 * Unlike the trigram guard, this catches PARAPHRASES of known results. NOT receipt-binding.
 */
export declare function assessSemanticNovelty(query: string, corpus: ReadonlyArray<ConjecturePriorArtEntry>, threshold?: number): Promise<SemanticNoveltyAssessment>;
//# sourceMappingURL=SemanticNoveltyEncoder.d.ts.map