Statistical uncertainty and clean-task non-inferiority

Determine the statistical precision and significance of the Vision-Language-Action recovery comparisons and establish whether recovery-data adaptation is non-inferior on clean-task performance.

Background

The reported Protocol A and Protocol B results use multiple training seeds but do not report per-seed dispersion. Consequently, the significance and precision of the observed recovery improvements cannot yet be assessed, and the reduction in clean-task success associated with targeted recovery-data training has not been shown to satisfy a non-inferiority criterion. Any appropriate uncertainty analysis must account for both independently trained policies and the paired structure of evaluation states and source episodes.