Making the Schedule an Input: Bitwise Reproducibility in Asynchronous ML Pipelines

Open-sourcing project code to ensure reproducibility is now common practice across the machine-learning community, but readers cannot necessarily use that code to reproduce a paper’s metrics exactly. Asynchronous pipelines can diverge across repeated runs even with fixed seeds and deterministic kernels, because runtime scheduling determines which version of an evolving state each component observes. We treat this otherwise hidden schedule as an experimental input, expose it as a logical schedule $S$, and fix $S$ using a reusable logical exchange contract. Across three systems, two independently developed Gaussian-splatting SLAM pipelines and an actor/learner RL trainer, three repeated runs at every declared working point were bitwise identical in both event history and application state. This held for a full 693-frame SLAM sequence and for 40,000 environment steps, including runs under two different declared schedules. A larger native-control experiment showed the scale of the uncontrolled variation: across 48 free-running runs spanning three seeds and sixteen repetitions per seed, all 48 produced distinct trajectories. Under the contract, repeated runs with the same declared inputs instead reproduced exactly. In one SLAM system, this reproducibility did not come at an apparent performance or runtime cost. All contract-controlled runs remained within the native performance range, at $0.84\times$--$1.09\times$ its center, and do not increase the actual running time. The contract fixes the input and output mapping of an asynchronous system, allowing experimenters to reliably separate the effects of environmental inputs and irrelevant parameters from those of declared inputs.