Agentic Language Models: A Proposal — Contract-Native Models, and a Governance Protocol for the Agentic Work That Comes Next

We propose two things. A model class: an Agentic Language Model is a language model whose native interface is the governance contract of the runtime it serves — a compiled, budgeted, digest-verified capsule in, a typed protocol act under constrained decoding out, a training signal derived from the runtime's governance gates rather than preference labels, and authority enforced externally by gates the policy cannot bypass. And where governance belongs: authority, budget, approval and evidence are substrate concerns, not application concerns, with GRIP offered as the open instance to build on. The paper gives the formal setting, membership criteria with a measurement procedure for each, and a public evaluation standard — a contract-fidelity suite built on counterfactual refusal injection, with thresholds pre-registered before any candidate's numbers exist. The harness code is not yet released, and no ALM claim is admissible until it is. The term has prior uses, which the paper prints; no priority of any kind is claimed.