Computational reproducibility

Reproducibility is the ability of someone else - including the author in six months - to regenerate the results from the code and data. It is not one practice but a stack: pinned environments (rseng-reproducible-environments), scripted pipelines (rseng-workflows), versioned data (rseng-data-management) and controlled randomness, assembled so that ONE documented command rebuilds the results. This skill owns the assembly and its verification; the layers below have their own skills. The bar to aim for: a stranger with the repository and the README reproduces the paper's numbers without emailing anyone.

The one-command bar

• Everything scripted, nothing manual: any step a human performs by hand (a click, a copy, an "then edit line 12") is a reproduction failure waiting to happen. Encode the full path from raw data to final figures/tables in a workflow or top-level script (rseng-workflows).
• One entry point, documented: make reproduce, snakemake all or ./run.sh - named in the README with expected runtime and resource needs. Long-running steps get cached intermediates so partial reruns are practical.
• Outputs land in generated directories, mapped to the paper: which script makes Figure 3 must be answerable from the repo (a results/README or a figures manifest).
• Configuration explicit: every parameter that shaped the published results lives in versioned config files, not command-line lore or notebook cell edits.

Determinism and honest nondeterminism

• Seed discipline: explicit RNG objects seeded from configuration, never global unseeded randomness; derive per-worker streams from a master seed for parallel runs; record seeds with outputs - a seed is provenance.
• Know the nondeterminism you cannot remove: thread scheduling, parallel reduction order, GPU kernels and library versions legitimately perturb low-order bits (rseng-numerical-accuracy). State the expected variability ("results match to 1e-6; figures identical") instead of claiming bit-identity you have not tested.
• Sort the unordered: filesystem listings, dict/set iteration and parallel completion order differ across runs; sort before anything result-bearing.

The research compendium

Structure the repository as a compendium - the recognized shape for reproducible research projects: data (raw read-only, processed generated), code, environment specification, outputs, and a README tying them together with the one command. Conventions and examples live at research-compendium.science. For projects headed to review, the compendium IS the replication package.

Replication packages and artifact evaluation

When results support a paper:

• Assemble the package: frozen code version (tagged release - rseng-publishing-releasing), data or scripted data retrieval with checksums (rseng-data-management), pinned environment, run instructions with runtimes, and a manifest mapping outputs to paper claims.
• Deposit, do not just link: an archival repository with a DOI (Zenodo-class) is the durable home; a git URL alone does not meet artifact-availability bars (ACM's artifact badging explicitly requires archival deposit for its Available badge).
• Target the venue's checklist when one exists (artifact evaluation tracks, journal data editors); rseng-software-peer-review covers review-side mechanics and CODECHECK-style independent execution.
• Declare AI involvement in producing the results in aidecl.yaml (rseng-ai-declaration) - reproducibility and provenance are the same promise at different layers.

Binder: reproducibility others can click

repo2docker builds a runnable image from a repository's standard environment files; mybinder.org hosts it so anyone can run the analysis in a browser without installing anything. Make a repo Binder-ready by keeping environment files canonical (no requirements drift), test the build locally with repo2docker before adding the badge, and expect image builds to rot as dependencies move - pin versions and re-test at releases (rseng-reproducible-environments). For compute-heavy work, Binder demos a subset; the full run documents its HPC path (rseng-hpc-computing).

Verify before you claim

Reproducibility untested is reproducibility absent:

• Clean-room test: fresh clone on a machine (or container) that never ran the project, follow only the README, compare outputs to the published ones with stated tolerances. This finds the undeclared dependency and the hardcoded path every time.
• Automate the claim where affordable: a CI job that runs the pipeline on reduced data and compares key numbers keeps the reproduction path from rotting between releases (rseng-ci-cd).
• Independent reruns (a colleague, a CODECHECK, a ReproHack-style event) are the strongest evidence - and normal practice, not an audit to fear.

Working with this skill

This skill is source-independent: its authority is the community reproducibility guidance and tooling linked below. It assembles what rseng-reproducible-environments, rseng-workflows and rseng-data-management provide layer by layer.

Learn more (verified):

• https://research-compendium.science - research compendium conventions and examples
• https://mybinder.org - Binder
• https://github.com/jupyterhub/repo2docker - repo2docker
• https://codecheck.org.uk/ - CODECHECK independent execution

Related skills

Check whether any of these applies before moving on:

• rseng-ai-declaration - declaring AI involvement in results
• rseng-archiving - depositing the package with a DOI
• rseng-data-management - versioned data with checksums
• rseng-numerical-accuracy - stating expected run-to-run variability
• rseng-publishing-releasing - tagged frozen release for the package
• rseng-software-peer-review - CODECHECK-style independent reruns