Code, data and results for: Reproducibility of Machine Learning Accuracy Claims for Chemically Stabilised Soils: An Audit of an Open Unconfined Compressive Strength Database

This archive accompanies the article "Reproducibility of Machine Learning Accuracy Claims for Chemically Stabilised Soils: An Audit of an Open Unconfined Compressive Strength Database" (submitted to Geotechnical and Geological Engineering). It contains the 194-record lime and cement stabilised soil database published by Azeem et al. (2026) under CC BY 4.0, the Python scripts used in the study, a table of every analysis setting (Table S1), and all numerical outputs behind the figures and tables. The scripts reproduce the 200-partition single-split experiment, the ten-times repeated five-fold nested cross-validation of five model configurations (AdaBoost, random forest, a regularised multilayer perceptron, extreme gradient boosting and monotonically constrained extreme gradient boosting), bootstrap and residual-corrected prediction intervals with empirical coverage, Shapley and ICE/PDP interpretation, the leverage-based applicability domain, and five software validation checks. Running run_all.sh (or run_all.bat) regenerates every figure, table and array in about 12 minutes on one CPU core; the archived outputs in results/ were regenerated this way and are identical to those reported in the paper.