Procedural generation with reproducible random streams

A seed is only part of the recipe

A world seed is useful only alongside the algorithm that consumes it. Adding one random draw can change every later result in a shared sequence. Reproducible generation therefore needs ownership of randomness, stable inputs, and a record of the generation rules, not merely a seed displayed in a menu.

Separate sequences deliberately

PCG documents repeatable initialization and distinct sequence selection in its random-generator interface. [1] NumPy documents deriving streams from root seeds and deterministic identifiers, including probabilistic independence qualifications. [2] Use a generator's supported stream or seed-derivation method; arbitrary arithmetic on neighboring seeds is not a substitute for a designed construction.

Worked example

An illustrative world uses seed 9137 and region (4, -2). Derive separate stream identities for terrain, loot, and decoration from a documented tuple: (generator_version, world_seed, region_id, subsystem). Specify its encoding and mixing algorithm so another machine reconstructs the same inputs.

Suppose a decoration update adds one extra flower-placement draw. With one global generator, that extra draw would shift the subsequent loot sequence whenever decoration runs first. With separate streams, it changes the decoration stream without advancing the loot stream. Terrain and loot remain reproducible provided their algorithms, inputs, and dependencies remain unchanged.

Next generate region (4, -2) before and after its neighbor. The intended result should match in both orders. If shared borders depend on neighboring output, resolve that dependency with an explicit rule; independent random streams alone do not make dependent generation order-independent.

Implementation workflow

Assign one owner to each stream and forbid accidental use of an unrelated global generator inside generation functions. Derive regional work from stable region identity, not worker-thread number or job completion order. Persist the seed, generator version, content version, and relevant generation settings.

For a partially generated world, save the required generator state or counters, or define a reproducible restart boundary. Record enough context in bug reports to regenerate the affected region. Keep fixtures that compare logical output rather than presentation noise.

Limits and verification

A library update, changed distribution sampler, reordered content catalog, or different floating-point decision can alter results despite an unchanged seed. Treat reproducibility as a versioned contract and preserve older generation logic or materialized content where existing saves require it.

Test changes in job scheduling, region traversal, and cosmetic settings. Check that weighted choices use the intended distribution rather than a biased hand-written mapping. Independent streams reduce accidental coupling; they do not guarantee a playable world, balanced rewards, or secrecy of generated outcomes.

Sources

[1] Using the minimal PCG implementation — Melissa E. O'Neill.

[2] Parallel random number generation — NumPy developers.