This replication study critically examines the computational methodology and findings of a seminal 2020 study that utilised machine learning to predict localised conflict onset in South Sudan. Using the original study's data sources, including ACLED and V-Dem, we re-implement the random forest classification model, systematically testing its sensitivity to feature engineering, temporal cross-validation, and class imbalance handling. Our analysis reveals significant variance in model performance upon adjusting the temporal granularity of the training data and the treatment of spatial autocorrelation, challenging the reported generalisability of the original predictive framework. The study underscores the critical importance of methodological transparency and robustness in computational conflict studies, offering revised protocols for future research in this domain.