We present BioModule V3, a lightweight transformer that predicts 17 biomechanical criteria per frame — including ground reaction forces, joint coordinates, torques, muscle activations, and neural excitations — directly from a 27-frame sliding window of 3D skeletal pose. The model is pre-trained on ground-truth (GT) pose inputs (GT Frozen) and then fine-tuned on top of seven state-of-the-art 3D pose estimators without updating their weights. Evaluated on Human 3.6M+ Subject 9 (Walking), BioModule V3 achieves low per-joint prediction error across all criteria while generalising across diverse pose estimator backbones. The per-joint architecture exposes spatially-resolved criterion timeseries, enabling downstream biomechanical analysis at the level of individual body segments.