export declare const normalizeLabel: (label: string) => readonly string[];
export declare const headTokens: (tokens: readonly string[]) => readonly string[];
export declare const similarity: (left: string, right: string) => number;
export interface IdentitySubject {
    readonly id: string;
    readonly kind: string;
    readonly labels: readonly string[];
    readonly owner?: string;
    readonly neighbours: ReadonlySet<string>;
    readonly distinctFrom: ReadonlySet<string>;
}
export interface NearDuplicatePair {
    readonly left: string;
    readonly right: string;
    readonly score: number;
    readonly corroboration: 'lexical' | 'owner' | 'neighbourhood';
}
export declare const strongLexicalThreshold = 0.95;
export declare const moderateLexicalThreshold = 0.8;
export declare const lexicalScore: (left: IdentitySubject, right: IdentitySubject) => number;
/**
 * Every candidate near-duplicate pair, sorted by subject id. Only subjects
 * of the same qualified kind are compared: kind is the cheapest structural
 * disagreement there is, and it also bounds the pairwise cost to the square
 * of the largest kind bucket rather than of the whole model.
 */
export declare const findNearDuplicates: (subjects: readonly IdentitySubject[]) => readonly NearDuplicatePair[];
/**
 * Both members of every candidate pair, mapped to their counterparts. The
 * question attaches to both sides deliberately: `ask --advise` filters open
 * questions to the subjects in a slice, so a finding recorded only against
 * the lexically-first member would vanish from a slice seeded on the other.
 */
export declare const nearDuplicateIndex: (subjects: readonly IdentitySubject[]) => ReadonlyMap<string, readonly string[]>;