Abstract

Backend substitution in Python-native machine-learning workflows is attractive for integration and performance reasons, but it can silently alter model-selection outcomes even when data partitions, seeds, and metrics are held fixed. This study addresses that risk for J48-style decision trees by asking whether an established decision-tree backend can be reused within standard scikit-learn control flow without materially changing workflow-relevant outputs. We evaluate a versioned J48-targeting Python backend through differential validation against WEKA J48 under matched experimental conditions, combining controlled alignment checks, heterogeneous general-tabular comparisons, intrusion-detection benchmarks, and workflow-level validation inside standard estimator pipelines. The retained reference baseline reproduces the compared WEKA outputs exactly throughout the controlled campaign, across seven of the eight general tabular datasets, and in the base configuration of all five intrusion-detection datasets, with CreditApproval retained as an explicit documented boundary. In a six-dataset repeated-fit workflow study, the strict and accelerated wrappers select identical hyperparameters and preserve held-out accuracy and macro-F1, while total grid-search-plus-refit time decreases from 78.54 s to 19.47 s, corresponding to a $4.03 imes $ speedup. These results support bounded, workflow-stable reuse of the released backend under the documented conditions and provide a practical validation pattern for backend substitution against mature external references in workflow-native toolchains.