Methods for Formal Verification of Agent Skills: Three Layers Toward a Mechanically Checkable Capability-Containment Proof

In three linesFormal verification paper for LLM agent skills. Presents three composable methods: sound static capability-containment analysis via abstract interpretation, refinement type system for tool-call envelopes, and SMT-bounded model checking. Open-source JavaScript implementation (enclawed framework) with 53 unit tests and end-to-end CLI demo.

Summary generated by Claude — human-verified