4.6 Downstream Analysis

4.6.1 Start from finalized data

Downstream analysis starts only from stage/manifest.json finalized by a successful execution Job, or from a prepared table explicitly allowed by the documents. First check sample counts, groups, missingness, matrix orientation, and metadata keys. Then choose statistics, microbiome, metabolomics, or machine-learning analysis that the data support.

4.6.2 Python, R, and results

The AI may generate Python or R scripts for EDA, statistics, plotting, and reports, but each script is saved in reproducibility/ before it is run by an isolated container. Pin dependencies, random seeds, and plotting parameters. Reports distinguish biological from technical replicates and record effect sizes, confidence intervals, and multiple-testing choices. Figures and HTML/PDF reports must trace back to the input summary and manifest.

4.6.3 Container safety

Prepare the exact BioLine NAME:VERSION after checking its bilingual downstream summary and empty fixed-command config. Pull only that version, use network none, remove extra capabilities, and mount stage/ and task reproducibility read-only. A container cannot modify upstream products or read the authoring workspace. Commands, stdout/stderr, exit codes, generated scripts, outputs, and logs remain in the current Job workspace. Historical local images are migration inventory, not runtime choices.

4.6.4 Boundaries of conclusions

Methods must be supported by the data structure and documents; do not add unsupported tests or figures just to fill a report. When inputs are insufficient, a method is unsuitable, or results are incomplete, record the concrete failure or gap and stop. Empty figures, partial reports, and runs without audit records cannot be marked completed.