| name | reproducible-research |
| description | Make an analysis re-runnable by anyone: enforce notebook hygiene (top-to-bottom-clean, no out-of-order state), pin environments and dependencies in a lockfile, set and thread random seeds, version data and artifacts, track experiments (params + metrics + code/data version per run), and convert a run-once notebook into a scripted, deterministic pipeline. |
Reproducible Research
Pin everything that can drift
Pin the Python version and every dependency in a lockfile (requirements.txt pinned / poetry.lock / conda env / container). A floating >= reproduces today and breaks on the next release. Deterministic install is the floor.
Version the data, not just the name
"The data" must be a version, not a name. Snapshot or content-hash the exact input (DVC / immutable snapshot) so a result is recoverable; "the latest table" makes every past result unverifiable.
Seed every stochastic step
A single global seed is not enough — the split, the model, the framework, and parallel ops each leak nondeterminism if you don't thread the seed through. Hunt the named sources of nondeterminism (unset seed, floating version, un-versioned data, thread-count-dependent ops) rather than shrugging at "flaky."
Notebook is a draft; the pipeline is the deliverable
An exploratory notebook is scratch — out-of-order cells, hidden globals. Make it restart-and-run-all clean, or extract it to a scripted, deterministic pipeline. Don't confuse a run-once notebook with a result.
Track the run, not just the result
Log params, metrics, artifacts, and the code+data version per run (MLflow or equivalent). Metrics without the commit and the data hash mean you can compare scores but never recover why.
Output
A reproducibility deliverable: a pinned env, seeds set and threaded, versioned data, tracked runs, a clean notebook, and a scripted re-runnable pipeline. Hand production CI/CD and serving to ml-engineering; the analysis content stays with exploratory-data-scientist / feature-and-modeling-engineer.
