Data Sciences
An existing analysis produced three byte-identical output files across two fresh Python virtual environments.
The first task was to check whether a saved analysis could produce the same outputs on a second run, and measure the time and memory it used.
First pilot: repeatability
4 November 2026 · Local experiment · Passed within the scope below
The lab reran its existing receipt-judge analysis unchanged, using frozen code and inputs. The workload covers 317 saved trace items, 67 clusters, five evaluated systems and 2,000 bootstrap resamples with seed 7. It recalculates results from saved model verdicts; it does not run new model inference or independently validate the labels.
| Measurement | Run 1 | Run 2 |
|---|---|---|
| Elapsed wall time | 17.842 s | 17.729 s |
| Peak resident memory | 25.688 MiB | 25.594 MiB |
| Exit code | 0 | 0 |
| Third-party packages | 0 | 0 |
Method and evidence
- Each run used a newly created, no-pip virtual environment, a separate working directory and empty HOME and temporary directories. CPython 3.12.15 ran in isolated mode on the same arm64 Mac, using macOS 27.0.1.
- SHA-256 hashes recorded the frozen code and inputs. The source files still matched after both runs. No package downloads, model calls or credentials were needed.
- Python's monotonic performance counter measured elapsed time around the time-wrapped analysis process. macOS
/usr/bin/time -lpmeasured peak resident memory. Environment setup was outside the timed task. - The generated tables, metrics JSON and per-item JSONL matched exactly. Python hashes and byte comparisons were checked independently with
/usr/bin/cmp.
Evidence tier: high for these observed runs and their output agreement. The lab retains the input manifest, runtime inventories, raw timing logs and comparison records. The underlying experiment traces are not published here.
What the result establishes
The analysis is repeatable under this frozen input and runtime combination. Two runs do not establish statistical performance stability, correctness of the research conclusions, or suitability for larger scientific workloads.
These were fresh virtual environments, not fresh machines or containers. They shared the interpreter installation, host, kernel and filesystem caches. Cache state and background host workload were not controlled. Peak memory was observed, not constrained by a hard memory cap.
Platform status and next scope
No data-science platform has been deployed. The proposed direction is version-locked execution images, marimo for Python/SQL exploration, and MLflow for experiment tracking. JupyterLab or JupyterHub remains an alternative for notebook interoperability, other kernels or multiuser administration.
Containerized runs, scientific-library workloads, cross-architecture compatibility, concurrency, hard resource limits, failure handling and human-facing entrypoints remain untested in this pilot. They remain in the next evaluation scope.