We participate in Track B of the SemEval 2026 Task 4 on narrative similarity, focusing on narrative representation learning. We introduce a synthetic dataset designed to disentangle core narrative aspects-abstract theme, course of action, and outcome-and propose a multi-head multi-positive extension of the InfoNCE objective to train aspect-specific embeddings. Our best model achieves 64.25\% accuracy on the test set. A nearest-centroid analysis indicates partial aspect-specific structure in the submitted checkpoint, while the training dynamics reveal a partial misalignment between the contrastive objective and the triplet-based evaluation protocol.