This technical report presents the approach and insights of team “Whatever” for the 2nd Edition of the WEAR Challenge. This challenge proposes an inertia-based Human Activity Recognition (HAR) problem focused on identifying sports activities using data collected by wearable sensors. To address this, we propose a multi-view ensemble model considering each view as the set of symmetrically worn sensors (i.e., right/left arm). The model is composed of an early encoder that shares weights across the views, followed by two view-specific branches (wrist and legs). We further encourage similarity between symmetric sensor positions by adding a similarity component to the loss function. We enrich inputs with frequency-domain and PCA features plus data augmentation. Our evaluation revealed that frequency-domain features and channel-wise random sign-flipping data augmentation were the main drivers of generalization, mirroring patterns seen on the challenge test set.