Researchers introduce atomic-quality probe for governing skill updates in compositional robot policies, addressing how library changes affect task success.

A paired-sampling cross-version swap protocol on robosuite manipulation tasks revealed a dominant-skill effect: on a dual-arm peg-in-hole task, one skill (ECM) achieved 86.7% atomic success rate while every other was at or below 26.7%, with the dominant ECM's presence in a composition shifting success rate by up to +50pp.

The atomic-quality probe (zero per-decision cost) combines per-skill probes with selective composition revalidation; on the T6 task, atomic-only scored 23pp below full revalidation (64.6% vs 87.5% oracle match), while a Hybrid Selector with m=10 closed most of that gap to ~12pp at 46% of full-revalidation cost.

Off-policy behavioral distance metrics failed to identify the dominant skill, ruling out this as a cheap predictor of composition outcomes when a skill in a robot's library is replaced.