A joint government–lab body gains power to block training or deployment

Where things stand (16 Oct 2026)

Nothing published since the 16 Sep 2026 opening date qualifies, and none of the structures now in play has the specific shape the question requires.

The government already exerts control, but only through the government.

• Executive Order 14409 (2 Jun 2026) set up a voluntary 30-day pre-release access framework. The NSA decides which models are covered, and the government and developer "collaborate" on choosing trusted partners. The order explicitly disclaims any licensing or preclearance (EO text).
• On 12 Jun the Commerce Department used export controls to force Anthropic to pull Fable 5 and Mythos 5. Access was restored after government approval (Anthropic, 12 Jun; Anthropic, 30 Jun).
• OpenAI limited the GPT-5.6 launch at the government's request.
• Lawfare calls this a "de facto licensing system" run by "various administration officials". The framework text is due to be released by 30 Oct under a FOIA settlement (Lawfare, 14 Oct).

All of this is either separate government and company review processes or government-only authority. Both are excluded.

What has happened inside the window is also excluded.

• White House Accord on Super Intelligence (29 Sep). It is self-policing: internal teams, outside auditors, and an independent committee of each company's board. No government seats and no binding power (text).
• Trump's floated committee. He mused about "a committee of sorts… maybe ten people… it could be from that group… so the committee can watch over the whole enterprise" (Mediaite).
• Super Intelligence Force. What followed instead was this task force, chaired by Director of National Intelligence Jay Clayton. Its members are officials plus outside advisers (David Sacks, Condoleezza Rice), not lab representatives. It has a 120-day report mandate, due around Feb 2027, and no binding authority (NPR; DW).
• Opposition to a formal regulator. Vice President Vance rejected an FDA- or FAA-style regulator (Tech Policy Press).
• Equity stakes. Trump floated "Intel-style" stakes in OpenAI and Anthropic. Those are passive, and equity without a joint body is excluded (Benzinga/Yahoo).

The governance models with momentum mostly fail the membership test.

• The idea of a self-regulatory body modelled on FINRA, the broker-dealer regulator, is the main one. Bessent drafted a proposal (Claims Journal/Bloomberg), and OpenAI, Anthropic and Google have discussed it (Forkast). These bodies normally have independent and industry board members and no sitting officials, and they are overseen by a separate agency. The Council on Foreign Relations expects Clayton to green-light an industry body with a separate government overseer (CFR).
• Democratic bills would create federal agencies, emergency kill-switches or licensed outside verifiers, all government-only or third-party (Transformer).
• Amodei's "pace the frontier" plan wants government to mediate antitrust problems, but says government "doesn't need to participate" (essay).

Base rate and paths to YES

Reference class. Formal US bodies that combine sitting federal officials and a private company, with binding power over the company's core product decisions, are nearly unheard of outside government corporations. Rescues, foreign-investment security agreements, the Intel stake and the US Steel golden share all avoided this structure. The base rate is close to zero, adjusted upward because this field is unusually crisis-driven: rogue-agent incidents, OpenAI's training pause, and chief executives calling for a slowdown.

Paths to YES, with rough probabilities through 2031:

• A crisis-driven joint oversight committee set up by contract, like the "Oversight Committee" in the AI 2027 scenario: about 4%.
• A self-regulatory or public-private body whose charter gives a federal official a seat, possibly non-voting and ex officio, which counts here. Precedent: NIST's foundation has its director as an ex officio member. About 3%.
• Trump's mixed committee, made binding through signed agreements: about 1.5%.
• Antitrust-sanctioned pacing coordination chaired by government, for example a Defense Production Act section 708 agreement: about 1%.
• A sitting official on a lab's board or safety committee: about 1%.
• Narrower joint committees for a single model, such as Mythos access: about 1%.

These add to roughly 11%. In the near term, the 30 Oct framework release is the only scheduled document that could surprise, but reporting describes a government-run process. I put the fourth quarter of 2026 at 0.6%.

Hazard shape

The quarterly hazard is about 0.6–0.75% through 2027, covering the task force report and any formalization, then 0.5% in 2028. It rises to 0.6% in 2029 under a new administration, then eases to about 0.45–0.5%, since a statutory government-only regulator would crowd out joint forms. I computed the cumulative series as 1 − ∏(1 − h) in Python. It gives 0.6% at end-2026, 3.2% at end-2027, 5.1% at end-2028 and 10.8% at end-2031.

Strongest case against my estimate

AI could become transformative within the window. Government–lab fusion is a widely discussed outcome, and the criteria are permissive: contractual authority counts, a single model counts, and non-voting ex officio seats count. A Trump-style "committee from that group", made binding by agreement, could appear quickly. I weighed this. It is offset by the need for both parties to publicly confirm the body, by its likely secrecy in a national-security setting, and by both parties' stated preference for government-only or industry-only models.

Question. By the last day of each quarter, have the US government and a covered developer announced a joint body, with sitting government officials among its members, that has authority to approve or block the training or deployment of that developer’s models?

Definitions.

• US government: the US federal government, acting through the White House or a federal department or agency; state and local governments do not count.
• Joint body: an identifiable committee, board, council, or other collective decision-making body formally established or designated by both the federal government and the covered developer, with members representing each. An existing body newly given the qualifying membership and powers counts. Separate company and government review processes do not constitute a joint body.
• Sitting government officials: at least one member currently holding federal public office or federal employment and serving on the body in an official government capacity. Former officials, private contractors, and officials serving solely in a personal capacity do not count. Official membership without a vote counts; attendance as an observer or invited adviser does not.
• Authority to approve or block: a law, order, executed agreement, corporate governance instrument, or adopted charter gives the body power to grant required permission for, refuse permission for, or order the suspension of qualifying training or deployment. Its decision must bind the developer under the governing arrangement, rather than merely recommend action to the developer or to a separate government decision-maker. Contractual or corporate-governance authority counts without a statutory regulatory power. No individual government member need have a veto, and no exercise of the power is required.
• Training or deployment: initiating, continuing, or resuming model training, including fine-tuning; releasing model weights; or putting a model into operational use internally or making it available to outside users. Authority over either training or deployment suffices, and it may apply to one identified model, a class of models, or all of the developer’s models. Authority confined to procurement, acceptance of a contract deliverable, or a government customer’s own use of a model does not count.
• Announced: both parties publicly confirm the same adopted arrangement, either in a jointly issued or jointly signed public document or in separate official publications. A statement of intent to negotiate or create such a body is insufficient. The body must have the qualifying authority and membership in effect by the qualifying confirmation; a charter assigning an occupied federal office an ex officio seat is sufficient evidence of government membership.

Resolves YES if qualifying public documentation establishes all of the above on or before the quarter-end. A published charter or contract may supply the membership and authority details omitted from the announcements. Both parties’ confirmations must fall within the Resolution window. If the requirements are established across several publications, the last necessary publication determines the resolution date.

Resolves NO if no qualifying announcement has been published by the quarter-end. Excluded are advisory panels, voluntary testing or model-access agreements without binding decision authority, government-only regulators, private-only safety committees, government equity or observer seats without the specified joint body and authority, and proposals awaiting adoption. The August 29, 2024 NIST agreements with Anthropic and OpenAI, described as model-access, research, evaluation, and feedback arrangements, do not themselves qualify. Nor does CAISI’s testing and collaborative-research remit itself establish a qualifying joint body.

Resolving source. Official White House or federal department or agency publications, including published orders, charters, and executed contracts, together with covered-developer publications on its own domain or developer-authored documents. A jointly signed document officially published by either party suffices for both confirmations. This is a narrower source requirement than Public confirmation (A2): third-party reporting alone cannot establish the required bilateral announcement. Apply Resolution date (A3) and the Resolution window.