Reproducible Research Compute Requires Environment Evidence
Holding
A computational result is reproducible only when its code, data reference, dependencies, parameters, and execution environment can be reconstructed.
Issue
Saving a notebook is not enough. Results may depend on package versions, hidden files, random seeds, credentials, hardware behavior, or mutable upstream data.
Analysis
The failure mode is an analysis artifact without an execution record. Another researcher can see the conclusion but cannot determine which environment produced it or whether a rerun used the same inputs.
Resolution
The system should capture immutable code revisions, dataset versions, dependency locks, parameter sets, randomization controls, execution logs, and output checksums. Restricted data should be referenced by governed identifiers rather than copied into uncontrolled locations.
Evidence Package
The evidence should include run manifests, environment definitions, data-access records, pipeline logs, output hashes, and documented deviations from the original run.
Conclusion
Research compute should preserve enough environment evidence to distinguish a rerun from an approximation.