Training-free model Jev falls far short of supervised HAR on accelerometer datasets
The paper (arXiv:2609.36154v1) evaluates Jev, a fixed general-purpose probabilistic decision model with no labeled examples or task-specific updates, on 1,800 class-balanced accelerometer windows from WISDM, UCI341, and PAMAP2. Jev's best deterministic representation produced macro-F1 scores of 0.038, 0.118, and 0.089 on those datasets respectively, versus 0.686–0.907 for three supervised HAR models; adding more numerical features worsened performance while a deterministic semantic rendering partially recovered it, and Jev's output probabilities were not reliably calibrated for recognition.