Oak & SparrowVolume I
Codifying Law
An agent-native reference for legal authority, deontic reasoning, provenance, jurisdiction and authorization boundaries.
Gatekeeper · Runtime AI governance · Oak & Sparrow
Deterministic pre-execution authorization for autonomous AI agents
AI agents can generate actions. Gatekeeper's job is to determine whether an actor has current, specific authority to execute an action before it crosses a controlled execution boundary.
Identity proves who is acting. Competence suggests what they can do. Neither grants authority to cause a specific consequence. Gatekeeper separates those questions using independently evaluated policy, evidence, delegation scope and action-bound authorization.
Governance before execution
Prompt instructions, agent self-reports and post-execution monitoring cannot establish that a particular operation remains authorized at the moment of consequence. Gatekeeper's pre-execution boundary is designed to evaluate the proposed operation and the evidence that supports it outside the model's own assertions.
Evaluate authority
Bind the actor, requested operation, target, policy context and current supporting evidence. An authenticated identity alone does not satisfy this determination.
Constrain delegation
Prevent delegated authority from silently expanding beyond its scope or persisting after the mandate or evidence supporting it has been revoked.
Record the outcome
Use controlled execution boundaries and evidence receipts so an authorization decision can be distinguished from the action that actually occurred.
How the enforcement architecture works
The architecture separates policy determination from action execution. A first decision gate evaluates whether an operation is admissible under applicable authority and evidence. A second, action-bound check and controlled actuator path constrain what can actually execute, with records for review. When authority or evidence is insufficient, the action is held or blocked under the applicable runtime policy.
Gatekeeper is built for consequential actions by humans, automation and autonomous agents. The relevant question is not whether an action was suggested by a model or a person. It is whether the specific consequence is authorized now.
Research language is not the runtime API
Codifying Law uses the deontic vocabulary ALLOW, TRANSFORM, HOLD and DENY to teach how legal propositions can be represented. Gatekeeper's public engineering descriptions discuss concrete enforcement verdicts such as ALLOW, HOLD and BLOCK. The theoretical vocabulary does not itself confer operational permission, and the two should not be mistaken for identical interfaces.
Engineering evidence and current qualification boundary
Oak & Sparrow has performed bounded Gatekeeper V2 policy-estate and controlled-route qualification work, including checks around authorization, evidence and execution adapters. Those results should be assessed within their tested scope. They do not, on their own, certify every external integration or qualify unrestricted production cutover.
The free public AI readiness scan examines publicly available evidence and is not Gatekeeper enforcement. Standalone Policy Packs are a separate licensed rule-software product. Reading this page or the research books does not grant access to protected execution capabilities.
The published research behind the architecture
Joshua Johosky's two public volumes explain how source-bound law, persistent actor identity, constitutional constraints and execution authority can be kept distinct. They are free research references, not live enforcement systems.
Volume II
Sentient Infrastructure Protocol
A constitutional framework for governed artificial agency, persistent identity, continuity and institutional accountability.
Explore the Governed Agency Foundations and author credentials.
Evaluate Gatekeeper against your action boundary
For enterprise security, governance teams, research analysts and technical evaluators, Oak & Sparrow can discuss the architecture, product boundaries, available test evidence and requirements for a scoped demonstration. No production authorization is implied by a briefing request.