Outputs and reproducibility

Each analysis writes to its configured experiment_dir. The exact set of files depends on the stages and task, but the run contains machine-readable evaluation artefacts together with figures and an HTML report.

Typical outputs

EXPERIMENT_DIR/
  configs.tsv
  configs.db
  tables/
  figures/
  ensembling/
  explainability/
  robustness/
  inference/
  report.html

Important artefacts include candidate metrics, out-of-fold predictions, MPMA-B selection outputs, ensemble selection metadata, method-specific explainability tables, robustness summaries, and report figures. Optional inference and experiment outputs depend on the configured stages. The inference workflow can reuse selected fitted preprocessing and learner pipelines; see Inference.

Re-running stages

Runs are designed to be restartable. To regenerate only the report:

mllabiome config.py --stage report

To regenerate explainability from existing upstream results:

mllabiome config.py --stage explain

A changed scientific configuration can alter run identity and should be treated as a new analysis unless the package explicitly recognizes the change as compatible with incremental evaluation.

Reproducible release record

For a manuscript or archived analysis, retain:

• the exact mllabiome release or Git commit;
• the complete Python configuration used for the run;
• input dataset versions and acquisition dates where relevant;
• random seeds and evaluation protocol;
• the environment lockfile or equivalent dependency record;
• generated manifests, selected-unit metadata, and out-of-fold predictions;
• the final tables and figures derived from that frozen run. The package records software and experiment provenance in run artefacts. For archival publication, pair those artefacts with a tagged software release and an external archive such as Zenodo.

Version used by these docs

These pages describe mllabiome version 0.2.0. The repository examples should be treated as the canonical source for the exact current configuration syntax.