Three-dimensional human pose estimation models are typ- ically evaluated using Mean Per Joint Position Error (MPJPE), a metric that measures positional accuracy of predicted joint locations. However, biomechanical applica- tions require joint angles such as flexion, abduction, and rotation, which are the clinically meaningful representation of human movement. We evaluate three state-of-the-art 3D pose estimation models (MotionBERT, VideoPose3D, Sim- pleBaseline3D) on the AthletePose3D dataset, a real-world athletic motion capture dataset, and compute ISB-compliant joint angles from predicted poses via a learned pose-to- angle mapping. Our key finding reveals a paradox: models with lower MPJPE produce higher joint angle error. Mo- tionBERT, the best model by MPJPE (37.2 mm), yields the worst angle accuracy (25.19°), while SimpleBaseline3D, the worst by MPJPE (58.5 mm), achieves the best angle accu- racy (19.21°). These results demonstrate that MPJPE alone is insufficient for evaluating biomechanical fidelity and argue for adopting angle-based evaluation metrics.