Predicted merger scores versus true scores for three test sets: on past-merger (left), future-merger (middle), and full-merger (right) scores. The red dashed line indicates the ideal case of perfect prediction. Each panel also displays the R² score for the corresponding test. Points are slightly transparent to indicate density. The model shows consistent predictive performance across different temporal subsets. High true scores tend to be underpredicted. This is mainly because the strongly skewed true-score distribution (Fig. 7): most halos have low scores, which dominate the training and bias the predictions toward low values.