Deterministic Simulation Seeds
How Kronos manages random-number seeds so stochastic Monte Carlo and training runs remain exactly reproducible.
Randomness must be controlled, not free
Monte Carlo transport and neural-network training are both stochastic: they consume streams of random numbers. Left uncontrolled, that randomness makes a result impossible to reproduce. Kronos manages seeds deterministically so that a stochastic run, despite being random, produces the same result whenever it is re-executed with the same seed and environment.
Seeds as recorded inputs
A seed is treated as a first-class input, recorded in the run manifest alongside code and data versions. Given the seed and the pinned environment, the pseudo-random stream is fully determined, so the run is reproducible. This is what lets a certified neutronics result or a trained surrogate be re-derived exactly rather than merely re-approximated.
- Seeds recorded in every run manifest
- Independent streams per parallel rank
- Reproducible parallel reductions of stochastic tallies
- Distinct seeds for genuine statistical replicas
Parallelism complicates seeding
In a parallel Monte Carlo run, each rank needs its own independent stream, or ranks would trace identical histories and the statistics would be wrong. Kronos assigns per-rank streams derived deterministically from a master seed, so the whole parallel run is both correctly independent across ranks and exactly reproducible as a whole.
Reproducibility also requires deterministic combination. Summing stochastic tallies across ranks in a floating-point-sensitive order can perturb the last bits, so certified runs fix the reduction order. This connects seed management to the broader precision strategy: controlled randomness plus controlled arithmetic yields a repeatable answer.
Seeds are not only for reproducibility; they are for honest statistics. Uncertainty studies deliberately vary the seed to generate independent replicas and estimate the Monte Carlo error, while keeping every replica individually reproducible. Deterministic seeding thus supports both exact replay and rigorous uncertainty quantification for both machines.