/**
 * Types for the pure-TypeScript ML kernel.
 *
 * Phase 1 of "self-learning Nexus": learn representations from the owner's own
 * accumulated observation corpus instead of hand-tuned heuristics. All artifacts
 * are plain JSON (Float32 matrices base64-encoded) persisted next to the other
 * ~/.nexus state, and all *inference* is pure typed-array math — no external
 * runtime, no API. Everything here is deterministic given a seed so the eval
 * gate that promotes a model is reproducible.
 */
/** Provenance + config recorded with every trained artifact. */
export type ModelManifest = {
    name: string;
    format: "nexus-sgns";
    version: number;
    dim: number;
    vocabSize: number;
    /** Hash of the training corpus — lets a loader know if the model is stale. */
    corpusHash: string;
    /** ISO timestamp, supplied by the caller (the kernel never reads the clock). */
    trainedAt: string;
    epochs: number;
    window: number;
    minCount: number;
    negatives: number;
    seed: number;
    /** Eval results from the gate that promoted this artifact (e.g. recall@10). */
    metrics?: Record<string, number>;
};
/** On-disk embedding artifact (plain JSON). */
export type EmbeddingArtifact = {
    manifest: ModelManifest;
    /** index → token. */
    vocab: string[];
    /** Row-major vocab×dim Float32 matrix, base64 of the little-endian buffer. */
    vectorsB64: string;
    /** Precomputed top-K neighbour indices per token, for O(1) getRelated(). */
    neighbors: number[][];
    /** Cosine score parallel to `neighbors`. */
    neighborScores: number[][];
    /** Subword feature vocabulary + matrix (present iff trained with subwords). */
    subwordVocab?: string[];
    subwordVectorsB64?: string;
    subwordCounts?: number[];
    minN?: number;
    maxN?: number;
};
export type Word2VecOptions = {
    dim: number;
    window: number;
    minCount: number;
    negatives: number;
    epochs: number;
    /** Subsampling threshold for frequent tokens (word2vec `t`); 0 disables. */
    sampleT: number;
    /** Initial learning rate (decays linearly toward 0). */
    alpha: number;
    seed: number;
    /** Neighbours to precompute per token for the query-time cache. */
    topK: number;
    /** Enable fastText-style subword (char n-gram) features. */
    subword: boolean;
    /** Subword n-gram range (over jamo-decomposed, boundary-marked words). */
    minN: number;
    maxN: number;
    /** Drop subwords whose total weighted count is below this. */
    subwordMinCount: number;
};
export type TrainedEmbeddings = {
    vocab: string[];
    dim: number;
    /** Row-major vocab×dim — the per-word input vectors. */
    vectors: Float32Array;
    /** Present when trained with subwords: the subword feature vocabulary and its
     *  row-major matrix, used to compose vectors for rare / unseen words. */
    subwordVocab?: string[];
    subwordVectors?: Float32Array;
    /** Per-subword backing count (how many in-vocab words contain it). */
    subwordCounts?: number[];
    minN?: number;
    maxN?: number;
};