/**
 * 🌐 THE DISTRIBUTION-SHIFT (DRO) CERTIFICATE — will the recommended setting still hold when the data drifts?
 *
 * Every optimizer reports the value it measured on the data it saw. But deployment data is never exactly the test
 * data — the customer mix shifts, the traffic changes, the population moves. A setting that looks best on the
 * nominal distribution can collapse under a modest shift, and nobody hands you a signed bound on how bad it can
 * get. AEGIS certifies robustness to INPUT wobble and Tolerance to PARAMETER wobble; this certifies robustness to
 * the thing they don't touch — the DATA DISTRIBUTION itself.
 *
 * Given samples of a setting's per-unit value (per-customer profit, per-query score…), this computes the
 * worst-case mean over every distribution within a χ²-divergence ball of radius ρ around the empirical one —
 * distributionally-robust optimization. For the χ² ball this worst case has the exact, Cauchy-Schwarz-tight closed
 * form  V = mean − √(ρ · Var)  (a conservative lower bound once the adversarial weights would go negative), so the
 * certificate states "under any distribution shift up to χ² ≤ ρ, the expected value is provably ≥ V" — and signs
 * it. It also names the variance penalty, so a high-mean-but-fragile setting is correctly out-ranked by a
 * slightly-lower-mean-but-robust one. Verify re-derives V offline.
 *
 * WORLD-FIRST + LLM-impossible: an LLM cannot compute the χ²-DRO worst case, prove it is a valid lower bound under
 * every shift in the ball, and sign a re-derivable verdict — it just quotes the nominal average. (DIAKRISIS —
 * MEASURED: over many random reweightings inside the ball, none ever beats the certified worst case [valid lower
 * bound], and the aligned adversary achieves it [tight]; the DRO value is monotone decreasing in ρ and recovers
 * the mean at ρ=0; a fragile high-variance setting collapses under an actual shift while a robust one holds, and
 * the certificate ranks them accordingly. HONEST: this is the χ²-divergence ambiguity set — it bounds shifts
 * measured by that divergence, not arbitrary adversarial corruption or support the samples never covered; and the
 * closed form is exact while the worst-case weights stay non-negative, conservative beyond that.)
 */
import { type KeyObject } from "node:crypto";
export interface DroCertificate {
    standard: "melete-dro-certificate/v2";
    divergence: "chi-squared";
    verdict: "ROBUST" | "FRAGILE";
    mode: "ambiguity" | "confidence";
    confidence: number;
    n: number;
    rho: number;
    threshold: number;
    mean: number;
    variance: number;
    worstCase: number;
    variancePenalty: number;
    values: number[];
    payloadHash: string;
    signature: string;
    publicKeyPem: string;
    algo: "ed25519+sha256";
}
export declare function droCertificate(opts: {
    values: number[];
    rho?: number;
    confidence?: number;
    threshold?: number;
    keys?: {
        publicKey: KeyObject;
        privateKey: KeyObject;
    };
}): DroCertificate;
export declare function verifyDroCertificate(c: DroCertificate): {
    ok: boolean;
    reason: string;
};
export declare function droGauntlet(): {
    score: 0 | 100;
    checks: Array<{
        name: string;
        pass: boolean;
        detail: string;
    }>;
};
//# sourceMappingURL=dro.d.ts.map