Tag
This paper presents three composable methods—abstract interpretation, refinement types, and SMT-bounded model checking—to mechanically verify that an LLM-driven agent skill's behavior is contained within its declared capabilities, closing the gap to the formal verification level proposed in a companion paper.