Reproducible Analysis
Invariants
- Raw data are read-only; transformed data go to a separate location.
- Paths are project-relative in shared code and configurable at the boundary.
- Parameters and thresholds are centralized and recorded.
- Every stochastic operation has an explicit seed.
- Software, reference database, annotation, and model versions are captured.
- Each output records its inputs, command/config, creation time, and checksum.
Workflow
- Define inputs, expected outputs, parameters, environment, and resource requirements.
- Add preflight checks for files, schema, identifiers, sample counts, and free space.
- Run from a clean process in documented order.
- Capture stdout/stderr with bounded logs and preserve non-zero exits.
- Verify output schema, row/sample counts, hashes, and scientific invariants.
- Rerun from clean state; compare deterministic outputs exactly and stochastic outputs under declared tolerances.
- Produce a run manifest and disclose partial or failed steps.
Notebook state is not rerun evidence. A visible figure is not proof that the final code and source data agree. Use evidence-review for the final consistency check.