Seeds control streams, not every source of nondeterminism

Inventory every random generator and the code path that consumes it. Seeding the environment does not necessarily seed sampled actions from its action space, the policy framework, NumPy Generator instances, or another library. In the pinned Gymnasium contract, an integer passed to reset initializes or resets its RNG, while reset with seed None preserves an already initialized stream. Reusing the same seed at every episode can repeatedly revisit the same initial random sequence rather than provide independent episode randomization. Match the procedure to the intended experiment.

Derive streams from stable logical identities and purposes, such as training replicate, actor identity, environment instance, episode, and sampler role, using a documented mapping and supported seed range. A physical rank can change on restart, so it is not always the right logical identity. Check collisions or truncation in any finite seed mapping and avoid assuming distinct integers prove statistically independent streams. Record the generator algorithm and version. A schedule that changes how many random draws are consumed can alter future behavior even when the initial seeds remain identical.

GPU kernels, floating-point reduction order, library algorithms, asynchronous actor arrival, and environment simulation can introduce variation unrelated to the root seed. Deterministic-operation settings can constrain supported operations on a pinned stack, sometimes with a performance cost, but are not a universal cross-device or cross-release promise. Record hardware, framework, simulator, wrapper, and numerical settings. A fixed seed can still produce a different replay insertion order when actors race, which then changes later minibatches and updates even if each local random stream is individually repeatable.

A restart needs the relevant generator states and consumption positions, not merely the original seed. It may also need environment state, recurrent state, replay contents, sampler state, normalizers, and pending-fragment policy. If the environment cannot serialize a mid-episode state, define recovery at a reproducible episode boundary or acknowledge a changed continuation. Do not claim exact uninterrupted equivalence from loading weights and calling reset with the old root seed. Test the claimed restart scope explicitly with tagged transitions and a short deterministic fixture.

Separate variability from training seeds from variability of evaluation episodes for one trained checkpoint. Many evaluation episodes reduce uncertainty about that checkpoint’s performance, but they do not replace independent training replicates when claiming an algorithm is reliable. Preserve all scheduled runs, including failures, and specify whether hyperparameters or checkpoint selection used the evaluation data. A single lucky seed can be exactly reproducible and still be unrepresentative. Publish the protocol and distributional evidence alongside the artifact needed to rerun individual cases.

The modular lab illustrates a finite mapping collision, not a random generator. Reset, explicitly set modulus eight and additive step two, predict the orbit period, then Play or Step. Compare a fresh Reset with modulus eight and step one explicitly set. Periods are four and eight; a deterministic stride can revisit a residue before covering all slots. This toy supplies no seed algorithm, independence test, environment dynamics, or framework determinism. It only makes the need to inspect a finite identity-to-stream mapping concrete.