Random seeds are only part of reproducibility
Keep numerical environment and state alongside a seed.
Research note · November 14, 2026Research question
Why can two runs with the same seed still differ?
Method
- Record seeds for each random-number generator and identify stochastic components. Preserve simulator, library, device and solver versions rather than treating the seed as a complete run identity.
- Repeat the same scenario in one environment, then in a deliberately changed environment. Compare trajectories within a declared tolerance and retain exact differences when they matter.
- Define the reproducibility claim: identical bytes, numerically close traces or equivalent task outcomes. These are different standards and should not share an unqualified reproducible label.
Records to retain
- Generator seeds and algorithms
- Software and device versions
- Initial state
- Comparison tolerance and outcome
Source basis
PyTorch explicitly notes that complete reproducibility is not guaranteed across releases, platforms or CPU and GPU execution. MuJoCo state documentation likewise shows why initial state must be specified independently of randomness.
Interpretation & limits
A fixed seed can reproduce a narrow setup without establishing broad stability. This protocol supplies no measured cross-platform agreement for Imprint.
Next experiments
- Repeat a small scenario before adding randomized stress.
- Publish the exact reproducibility criterion used for comparison.
References
Technical reference
Implementation details supporting this note, drawn directly from the maintained documentation. These are technical references, not additional papers or new experimental results. Expand a topic to inspect its method, fields and limits.