When Validation Stops Learning: Auditing Update Admission for Continual Embodied Agents

Summary

This position paper argues that update admission in continual policy learning must be audited not only for error control but for the learning opportunities a gate destroys at a stated interaction budget. The central observation is that a range-based Hoeffding admission gate cannot certify unchanged old-task behavior within realistic budgets: at m=3 comparisons and delta=epsilon=0.05 it needs 4,229 pairs per comparison (25,374 episodes with equal allocation), exceeding a 20,000-episode cap, whereas a paired Clopper-Pearson construction exploiting rare outcome disagreements needs only 117 pairs. The authors add certified historical-reference promotion and a round-level missed-opportunity metric that does not count normal candidate selection as error. In a deliberately simple one-step pushing simulator (ten sequential tasks, 32 seeds), fresh paired checks admit 81/256 updates at a 2,000-episode stage cap versus zero for Hoeffding and empirical Bernstein, with no observed contract violations; at 20,000 episodes the paired gate admits 229/256. Closed-loop, however, the paired gate still misses 60.4% of opportunities and unguarded replay attains higher final success (100.0% vs 59.6%), so the audit documents gating's opportunity cost rather than a better learner. A separate factorial stress test shows fresh proxy feedback does not remove selection optimism (model gap 41.9 vs 40.8 percentage points at curvature 0.35). Main limitations: all evidence is synthetic scalar simulation; no LIBERO, physical-robot, or VLA experiments; code and raw outputs are excluded from the preprint package.