The White House AI framework is done. Nobody can read it.

The White House AI framework for reviewing frontier models hit its statutory deadline on August 1. That's the entire public record. The administration confirmed the document exists and is complete, then declined to say what's in it, who has seen it, or when the labs start following it. If your business depends on a model shipping when the vendor says it will, a process you cannot read now sits between you and that date.

What actually happened

Executive Order 14409, signed June 2, gave agencies 60 days to stand up a voluntary early-access framework for "covered frontier models." Per Axios, a White House official confirmed the framework "was complete by the deadline" — and would not confirm whether it will ever be published. CNBC reports the administration is hosting the major labs to review the finished framework with the Office of the National Cyber Director. Google, OpenAI, and Anthropic reviewed an earlier draft together and filed edits in late July.

The mechanism itself is public, because the EO is. Per Norton Rose Fulbright's analysis, developers can voluntarily give the federal government up to 30 days of access to a model before it goes to other trusted partners, with the government and the developer jointly picking who else gets in early. The EO states plainly that this is not licensing, pre-clearance, or permitting. Which models qualify as "covered" gets decided through a classified benchmarking process for advanced cyber capabilities — thresholds undisclosed.

So: participation is voluntary, the criteria are classified, the framework is unpublished, and the review window is 30 days. Four unknowns stacked on top of your vendor's ship date.

Why frontier model gating matters for your business

We wrote in July about a FINRA-style AI watchdog when it was still a proposal. It's now a finished document being walked through in a closed meeting. The practical effect is the same one we flagged when GPT-5.6 shipped government-gated: model availability is a policy dial, and you don't have a hand on it.

Nobody is coming for your automation. But if you built a workflow that assumes a specific model at a specific version on a specific date — a launch tied to a release, a cost model tied to a new tier's pricing — you've made a bet on a schedule that a classified benchmark can move. Voluntary frameworks also don't stay voluntary in practice once every major lab signs on; they become the floor.

The defense is boring and it works: abstract the model layer. Route through a config value, not a hardcoded string. Keep a second provider warm with a real eval you run against both, not a hopeful comment in the README. Write your prompts against capabilities you can name — tool calling, a context length, a structured-output mode — rather than against one vendor's quirks. Then a delayed release is a config change, not a replan.

The tell that you're exposed: if a model your business runs on went dark for 30 days, could you name what breaks? If the answer takes more than a minute, that's the work.

Sources: Axios, CNBC, Norton Rose Fulbright on EO 14409.