Calibrated Does Not Mean Reliable: Lightweight Uncertainty for Molecular Trajectory Prediction

Uncertainty estimates could help determine when learned molecular predictions need further scrutiny, but interval coverage alone may be misleading. We study ATOM1.5: a 5,633-parameter Gaussian variance head attached to a frozen molecular trajectory operator, preserving its coordinate predictions exactly. In a matched three-molecule comparison, moving from random-frame to guarded temporal splits reduces error-uncertainty correlation from 0.216 to 0.0693. Across three seeds, validation rescaling reduces coverage error to 0.0269, yet selecting the lowest-uncertainty quarter increases error on ethanol. Molecule-shift detection and rotation robustness also remain inconsistent. These results distinguish empirical calibration from useful risk ranking, an important boundary before molecular uncertainty guides discovery decisions.