Reproducibility in computational biology depends on more than clean code, and machine learning (ML) workflows introduce failure points that classical bioinformatics pipelines never had to consider.