/**
 * FEDERATED META-BRAIN — the data flywheel that doesn't break sovereignty. The hard truth about a from-
 * scratch optimizer: every customer starts cold, so the vendor never gets smarter and there's no moat. But
 * you can't pool customers' raw data either — a pharma lab's secret recipe must never leave the building.
 *
 * The META-BRAIN squares that circle. After a run, each site shares only a privacy-safe LANDSCAPE
 * FINGERPRINT — abstract shape/behaviour features (how rugged, how smooth, how many effective dimensions,
 * the shape class, how concentrated the good region is) — and NEVER the optimum's coordinates or any
 * measured value. The central registry learns which SEARCH STRATEGY suits each class of landscape. A new
 * run whose fingerprint matches gets a recommended search profile (how much to explore vs exploit, how many
 * seeds) instead of a blind default — so it converges faster. The more sites contribute, the sharper the
 * strategy, and none of them leaked a secret.
 *
 * Honest by construction (DIAKRISIS): the fingerprint shares HOW a landscape behaves (to tune the search),
 * NOT WHERE the answer is (that would leak the recipe) — the gauntlet proves no raw value or coordinate
 * appears in what's shared. The benefit is MEASURED (the recommended profile beats the wrong one on that
 * class), not a promised "5 instead of 20" — the gain depends on how well a new problem matches known ones,
 * and is bounded. This is meta-learning of SEARCH STRATEGY, not a magic global oracle.
 */
import { type Space, type Experiment } from "./space.js";
import { type Observation, type Goal } from "./engine.js";
export interface LandscapeFingerprint {
    dims: number;
    ruggedness: number;
    smoothness: number;
    effectiveDimsFrac: number;
    concentration: number;
}
export interface SearchProfile {
    seeds: number;
    exploreWeight: number;
    note: string;
}
export interface MetaEntry {
    fingerprint: LandscapeFingerprint;
    bestExploreWeight: number;
    bestSeedsFrac: number;
    n: number;
}
/** A privacy-safe summary of HOW a landscape behaves — never WHERE the optimum is, never a raw value. */
export declare function landscapeFingerprint(obs: ReadonlyArray<Observation>, space: Space, goal?: Goal): LandscapeFingerprint;
/** Recommend a search profile for a new run, learned from prior landscapes of the SAME behavioural class. */
export declare function recommendProfile(fp: LandscapeFingerprint, registry?: ReadonlyArray<MetaEntry>): SearchProfile;
/** Append a landscape's learned strategy to the registry — anonymised (fingerprint only). CRDT-style merge. */
export declare function contributeFingerprint(registry: ReadonlyArray<MetaEntry>, fp: LandscapeFingerprint, bestExploreWeight: number, bestSeedsFrac: number): MetaEntry[];
/** A minimal profile-driven search used to MEASURE that a recommended profile actually helps. */
export declare function profiledSearch(space: Space, oracle: (e: Experiment) => number, budget: number, profile: SearchProfile, goal?: Goal, seed?: number): {
    best: Observation;
    evaluations: number;
};
export declare function metabrainGauntlet(): {
    score: 0 | 100;
    checks: Array<{
        name: string;
        pass: boolean;
        detail: string;
    }>;
};
//# sourceMappingURL=metabrain.d.ts.map