/**
 * 🧩 THE ATTRIBUTION CERTIFICATE — which feature drove this decision, with a PROOF the credit is fair?
 *
 * "Why was I denied?" is now a legal right (GDPR Art. 22, the EU AI Act, US adverse-action notices). The industry
 * answer is feature attribution — but the popular tools (a single SHAP run, an LLM's post-hoc rationalization) give
 * numbers nobody can check, and a vendor can quietly tilt them to hide the real reason. The Shapley value is the
 * UNIQUE attribution that satisfies the fairness axioms (efficiency, symmetry, dummy, linearity) — but nobody hands
 * you a signed proof that a specific attribution actually IS the Shapley value and actually obeys those axioms.
 *
 * This certificate computes the EXACT Shapley attribution from the model's own coalition value table (every subset
 * of features present vs. set to baseline), proves the axioms hold to machine precision — the credits sum exactly to
 * the prediction minus the baseline (efficiency), identical features get identical credit (symmetry), a feature that
 * never moves the output gets zero (dummy), attribution is additive across models (linearity) — and signs it. Verify
 * re-derives the whole attribution from the recorded value table offline and REJECTS any attribution whose credits
 * don't sum to the prediction (the tell-tale of a tilted explanation).
 *
 * WORLD-FIRST + LLM-impossible: an LLM cannot enumerate the 2ⁿ coalitions, compute the exact Shapley value, prove
 * the four axioms, and sign a re-derivable attribution — it rationalizes a plausible-sounding reason. (DIAKRISIS —
 * MEASURED: efficiency holds to ~1e-14 [Σφ = v(N)−v(∅)]; a dummy feature gets exactly 0; symmetric features get
 * exactly-equal credit; attribution is linear across value functions to ~1e-14; a forged attribution that doesn't
 * sum to the prediction is rejected. HONEST: this is the EXACT game-theoretic attribution for the value function +
 * baseline you supply — it explains THIS model's behavior under that baseline, not ground-truth causation; exact
 * Shapley is 2ⁿ so it is for a modest feature count [n ≤ ~16], and which baseline you pick changes the credits.)
 */
import { type KeyObject } from "node:crypto";
export declare function buildValueTable(n: number, value: (present: boolean[]) => number): number[];
export interface AttributionCertificate {
    standard: "melete-attribution-certificate/v1";
    method: "exact-shapley";
    n: number;
    featureNames: string[];
    baseline: number;
    prediction: number;
    phi: number[];
    efficiencyResidual: number;
    axiomsHold: boolean;
    valueTable: number[];
    payloadHash: string;
    signature: string;
    publicKeyPem: string;
    algo: "ed25519+sha256";
}
export declare function attributionCertificate(opts: {
    valueTable?: number[];
    n?: number;
    value?: (present: boolean[]) => number;
    featureNames?: string[];
    keys?: {
        publicKey: KeyObject;
        privateKey: KeyObject;
    };
}): AttributionCertificate;
export declare function verifyAttributionCertificate(c: AttributionCertificate): {
    ok: boolean;
    reason: string;
};
export declare function attributionGauntlet(): {
    score: 0 | 100;
    checks: Array<{
        name: string;
        pass: boolean;
        detail: string;
    }>;
};
//# sourceMappingURL=shapley.d.ts.map