The EO Didn't Fix the Captured Framework. It Became It.
The Trump administration's voluntary pre-release framework for frontier AI does not constrain the labs. It formalizes them as the architects of their own assurance layer.
The mechanics are straightforward. Under the framework, labs provide advance model access to federal agencies before broader release. Agencies run their own cybersecurity benchmarks. No independent body evaluates the substance of those assessments. Lab cooperation is not a condition the framework enforces: it is the mechanism the framework runs on. The evaluated party and the assurance layer are the same entity.
This is not a flaw the EO introduced. It is an architecture the labs had already built and exported before Washington arrived.
OpenAI's policy documents describe their government engagement as helping policymakers "understand, improve, and shape" AI policy. The direction of that influence runs outward from the labs. The "for Countries" commercial initiative translated that positioning into a product: sovereign AI infrastructure packages bundling capability access with governance tooling and access controls. Allied governments buying the capability received the governance vocabulary as part of the transaction. The assurance architecture was installed in partner national strategies before any federal framework existed.
The EO's contribution is not correction. It is adoption. The voluntary framework takes the structural arrangement the labs had already commercialized and ratifies it as federal policy. The labs remain architects of the framework rather than its subjects. Technical legitimacy stays concentrated where it has always been. The governance layer is real. The question of who authored its outer boundary is settled.
That is the condition the EO formalizes.