Architecture of the Step 2 cosmic-ray filter. To break the spatial degeneracy between cosmic-ray events and genuine sources, a dual-branch network integrates multimodal inputs. The temporal branch applies a transformer encoder with an adaptive sliding window to the three-channel light curve (time, rate, and error). Concurrently, the spectral branch processes the 1D pulse invariant (PI) via a multilayer perceptron (MLP). Globally pooled embeddings from both modalities are concatenated for the final classification.