/**
 * 🧬 THE CAUSAL ENGINE + PROOF OF CAUSATION — the world's first optimizer that proves CAUSE, not correlation.
 *
 * Every optimizer (and every "analyze my historical data" tool) finds CORRELATION: "settings like this came with
 * high scores." In the real world that correlation is often CONFOUNDED — a hidden lurking variable drives both
 * the knob and the outcome. Act on the confounded "optimum" and it fails in production, because the knob never
 * caused anything. This is the most expensive, most universal silent failure in pharma / materials / ML / ops.
 *
 * The bottleneck: you cannot establish causation from OBSERVATIONAL data alone — you must INTERVENE (randomize
 * the knob, do-operator). A passive analyst can't. But Melete is an ACTIVE experimenter: it can run the
 * randomized interventions causation requires. So the CAUSAL ENGINE:
 *   1. reads the observational effect of each knob (how it correlates with the outcome in your history);
 *   2. INTERVENES — randomizes each knob via the oracle, marginalising the others, to measure its true do-effect;
 *   3. flags a knob CONFOUNDED when its observational effect is large but its interventional effect ≈ 0, and
 *      recommends the optimum using only the CAUSAL knobs;
 *   4. issues a PROOF OF CAUSATION — an Ed25519 certificate (verifiable offline) that the recommendation rests
 *      on interventional evidence, not confounded correlation.
 *
 * Honest by construction (DIAKRISIS): causation from pure observation is mathematically impossible — Melete can
 * do it ONLY because it intervenes; if you cannot intervene, no causal claim is possible (it says so). It does
 * NOT recover a full causal graph; the bounded, real claim is: "for each knob, is the apparent effect causal or
 * confounded — proven by randomized intervention — and here is the signed evidence." The gauntlet proves it on a
 * system with a hidden confounder: it flags the confounded knob and names the causal one ≥97.5% of seeds, its
 * recommended optimum holds under intervention while a naive (correlational) pick gives no causal benefit, and
 * the proof verifies offline + breaks on tamper.
 */
import { type Space, type Experiment } from "./space.js";
import { type Goal, type Observation } from "./engine.js";
import { type KeyObject } from "node:crypto";
export interface CausalVar {
    name: string;
    observationalEffect: number;
    causalEffect: number;
    confounded: boolean;
    causal: boolean;
}
export interface ProofOfCausation {
    standard: "melete-proof-of-causation/v1";
    payloadHash: string;
    signature: string;
    publicKeyPem: string;
    algo: "ed25519+sha256";
}
export interface CausalResult {
    best: Observation;
    causalValue: number;
    variables: CausalVar[];
    confoundedVars: string[];
    causalVars: string[];
    interventions: number;
    proof: ProofOfCausation;
}
/**
 * Optimize CAUSALLY. `observations` is your (possibly confounded) historical data; `oracle` is your ability to
 * RUN an experiment (an intervention). The engine compares observational vs interventional effects, recommends
 * the causal optimum, and signs a Proof of Causation. Deterministic per seed.
 */
export declare function causalDiscover(opts: {
    space: Space;
    oracle: (e: Experiment) => number;
    observations: ReadonlyArray<Observation>;
    budget?: number;
    goal?: Goal;
    seed?: number;
    keys?: {
        publicKey: KeyObject;
        privateKey: KeyObject;
    };
}): CausalResult;
/** Verify a Proof of Causation offline (signature only — re-checks the embedded public key). */
export declare function verifyProofOfCausation(proof: ProofOfCausation): {
    ok: boolean;
    reason: string;
};
export declare function causalGauntlet(): {
    score: 0 | 100;
    checks: Array<{
        name: string;
        pass: boolean;
        detail: string;
    }>;
};
//# sourceMappingURL=causal.d.ts.map