/**
 * BATCH PLANNER — the physical world runs experiments in PARALLEL. A lab has eight reactors, a fab has eight
 * print heads, a greenhouse has eight plots, a cloud sweep has eight GPUs. A one-at-a-time optimizer wastes
 * them: it proposes a single next experiment and makes the other seven sit idle until the first comes back.
 * BATCH PLANNER proposes the k MOST VALUABLE experiments to run together — and, crucially, makes them
 * DIVERSE, so you aren't burning all eight machines on eight near-identical settings.
 *
 * It scores every candidate by optimistic potential (a Lipschitz upper bound from your data), then picks
 * greedily with a diversity penalty: once a setting is chosen, everything near it loses appeal, so the next
 * pick jumps to a different promising region. The result is a spread of high-potential experiments that
 * explore complementary parts of the space at once — same total experiments, k× fewer rounds of waiting.
 *
 * Honest by construction (DIAKRISIS): the value is wall-clock (you finish in fewer ROUNDS because the
 * machines run in parallel), not fewer total experiments; a purely sequential optimizer can use the
 * information between each run, which a batch cannot. BATCH PLANNER is for when you HAVE parallel capacity.
 * The gauntlet proves the batch is genuinely diverse — it covers separate optima instead of clustering.
 */
import { type Space, type Experiment } from "./space.js";
import { type Observation, type Goal } from "./engine.js";
/** Propose k diverse, high-potential experiments to run in PARALLEL this round. */
export declare function proposeBatch(space: Space, obs: ReadonlyArray<Observation>, goal: Goal, k?: number, seed?: number): Experiment[];
export declare function batchGauntlet(): {
    score: 0 | 100;
    checks: Array<{
        name: string;
        pass: boolean;
        detail: string;
    }>;
};
//# sourceMappingURL=batch.d.ts.map