/** * RELIABILITY — the falsifiable "97.5% on every landscape" bar. * * A single benchmark function proves nothing; an optimizer that wins on a smooth bowl can fail badly on a * deceptive, multi-trap, or high-dimensional surface. RELIABILITY runs the engine across a deliberately * adversarial BATTERY — smooth, many-trap (Rastrigin), deceptive (Ackley), curved-valley (Rosenbrock), * oscillatory (Griewank), high-dimensional, and a sharp needle — each NORMALISED so the true global * optimum is exactly 1.0. The score for a run is therefore literally "% of the true optimum reached". * * The bar: the engine must reach a high fraction of the true optimum on EVERY landscape, averaged over * several seeds — not just on average across the battery (which would let a smooth-surface win hide a * pathological failure). The gauntlet reports the worst landscape, so the claim is honest and falsifiable: * anyone can re-run it and check the per-landscape numbers. */ import { type Space, type Experiment } from "./space.js"; import { type Goal, type Observation } from "./engine.js"; export interface Landscape { name: string; space: Space; f: (e: Experiment) => number; budget: number; note: string; } /** * POLISH — a deterministic compass / pattern search (Hooke–Jeeves style) that follows a curved valley the * global explorer can only get near. From the incumbent best it probes ± a step along each axis; on an * improvement it keeps the move, otherwise it halves the step — so it converges into ridges (Rosenbrock) * that random/Bayesian sampling burns hundreds of experiments failing to nail. No randomness, no deps: * the local exploiter that turns "near the optimum" into "on the optimum". */ export declare function polish(space: Space, oracle: (e: Experiment) => number, start: Experiment, evals: number, goal?: Goal): { experiment: Experiment; value: number; }; /** * RELIABLE DISCOVER — Melete's signature 3-paradigm memetic engine. Three fundamentally different searchers * compose into one run: (1) the self-allocating PORTFOLIO explores globally to find the right basin; * (2) an OPTIMISTIC LIPSCHITZ INFILL (DIRECT-style) turns Melete's own optimality certificate INTO an * acquisition — it samples the single point where the certified upper bound is highest, i.e. the most * promising unexplored region the data cannot yet rule out; (3) NELDER–MEAD polish exploits locally to nail * the optimum. Global basin → certificate-guided infill of the gaps → local precision. Fully deterministic * ⇒ reproducible + signable. The diamond (CERTIFY) is not just a report here — it steers the search. */ export declare function reliableDiscover(opts: { space: Space; oracle: (e: Experiment) => number; budget: number; seed?: number; goal?: Goal; }): Promise<{ best: Observation; armStats: import("./portfolio.js").ArmStat[]; evaluations: number; goal: Goal; }>; /** All landscapes are maximise, normalised so the global optimum value == 1.0 (so best.value == % of optimum). */ export declare function landscapes(): Landscape[]; export interface LandscapeResult { name: string; meanPct: number; minPct: number; note: string; budget: number; seeds: number; } /** Run the portfolio engine over every landscape × `seeds` seeds; return % of true optimum reached. */ export declare function reliabilityBench(seeds?: number): Promise<{ results: LandscapeResult[]; worst: LandscapeResult; }>; export declare function reliabilityGauntlet(): Promise<{ score: 0 | 100; checks: Array<{ name: string; pass: boolean; detail: string; }>; }>; //# sourceMappingURL=reliability.d.ts.map