Safe Skill Retirement for Physical Agents

The paper proposes auditing skill retirement for physical agents with matched-authority counterfactuals, that is, holding the requested action, tool parameters, and expected effects fixed while varying a single governance predicate at a time, and uses a two-gate retirement certificate to require that a candidate reduction both preserves authorized utility and produces no unauthorized effects within the protected scope. Experiments cover four frontier and local model configurations and twelve skill packages totaling 2592 evaluation units; task-certified reduction removed over 94% of skill clauses while preserving authorized completion, but every skill package exhibited unauthorized protected effects. Boundary enforcement eliminated such effects in the declared audit but failed the utility gate on one configuration; a bounded compositional protocol passed both gates on all four configurations with a utility margin of zero. The authors also completed real-device end-to-end validation on a read-only Home Assistant camera link.