Sources of nondeterminism

Reproducibility can be lost in a few specific ways. This chapter lists them.

Seeds

Two runs diverge if their seeds differ. Conversely, if two runs with different seeds produce the same trajectory, the seed is not reaching the draw sites — check that the model is built with with_seed and that draws go through DetRng::from_context.

Thread count

A correct model is thread-count invariant. If a trajectory changes with the number of workers, some step depends on visitation order. The usual culprits are a shared mutable RNG, a floating-point sum accumulated in completion order, or a collected set whose order was not stabilised. See reproducibility.

Floating-point behaviour

Floating-point results are reproducible on the same target but are not guaranteed identical across:

• different CPU architectures or compilers,
• fused-multiply-add contraction settings, or
• different math-library versions.

Report the target and toolchain alongside numerical results, and compare trajectories on the same platform. Within one platform, order-independent accumulation keeps results exact.

External input and output

Anything the model reads from outside — a data file, the clock, the environment — is part of its inputs. A run is only reproducible if those inputs are fixed. Pin input data by content, avoid reading wall-clock time into model state, and record which inputs a run used.

GPU execution

GPU results depend on the adapter, its driver, and the shader compiler. A GPU run is reproducible on the same adapter and driver, but not necessarily across different hardware, and not necessarily bit-identical to the CPU path. Where cross-platform reproducibility matters, run on the CPU or fix the GPU environment, and validate the CPU/GPU equality tests on your hardware.

Summary