/**
 * rng.ts – shared, optionally-seeded random source for weight initialisation.
 *
 * Weight init across the model and every block used to duplicate the same
 * `Math.random()` Box–Muller draw. That made cold-start weights
 * non-reproducible across machines. This module centralises the draw and lets
 * the model install a deterministic seed for the duration of construction, so
 * the same `seed` yields byte-identical initial weights everywhere.
 *
 * The default (unseeded) source delegates to `Math.random`, preserving the
 * original behaviour for callers that don't request a seed.
 *
 * The seeded generator uses the same LCG constants as tools/generate-bin.js so
 * tooling and runtime agree on what a "seed N" model looks like.
 */
/** Deterministic linear-congruential generator (Numerical Recipes constants). */
export declare class SeededRng {
    private _s;
    constructor(seed: number);
    /** Next float in [0, 1). */
    next(): number;
}
/**
 * Installs (or clears) the deterministic init seed.
 * Pass a number to make subsequent `randn`/`gaussianArray` draws reproducible;
 * pass `undefined` to restore the default `Math.random` source.
 *
 * Construction is synchronous, so a process-wide source is safe: seed before
 * building a model and clear afterwards.
 */
export declare function setInitSeed(seed: number | undefined): void;
/** Box–Muller Gaussian sample from the active source. */
export declare function randn(std?: number): number;
/** Returns a Float32Array of `n` Gaussian samples with the given standard deviation. */
export declare function gaussianArray(n: number, std: number): Float32Array;
//# sourceMappingURL=rng.d.ts.map