Model performance as a function of τ, evaluated using three target score definitions: past-merger (left), future-merger (middle), and full-merger (right) scores. Orange and blue curves show the training and test R², respectively, and gray histograms indicate the sample count. This test uses the leakage-minimized split (grouping by halo ID only, so that progenitors and descendants of training halos are excluded from the test set) together with the rotation-based filtering scheme for sample selection.