The August 1 federal AI framework deadline came and went with nothing

August 1 was the deadline written into Executive Order 14409 for the federal frontier AI framework. It came and went. As of August 2 there were no Federal Register notices, no NIST or CISA publication, and no statement from OSTP. All three deliverables the order promised are missing: the classified benchmarking program (NSA, CISA, NIST), the voluntary frontier-disclosure framework (Treasury, NSA, CISA, NIST), and the cyber-workforce expansion plan (OPM). If you run models or agents in production and your compliance team asked what you had to do by August 1, the answer is: federally, nothing, because nothing exists.

What the framework was supposed to be

The mechanism at its center is pre-release evaluation through CAISI and the TRAINS program at Commerce, focused on national-security risk. Five labs confirmed participation: OpenAI, Anthropic, Google, Microsoft and xAI. Meta did not. In the days before the deadline, Sam Altman met White House chief of staff Susie Wiles, and on July 30 in the Senate he conceded that development might need to be 'paced'. That was the political high-water mark. The paperwork never followed.

Why Meta is out, and why that matters more than it looks

TRAINS is designed for closed-API models: a review window before release, during which a model can be evaluated and, in principle, held. Open weights do not fit that shape. Once Llama weights are on the internet, no review can pull them back. Meta's holdout is not obstinacy; it is a statement that the mechanism is architecturally incompatible with how Meta ships.

That has a direct consequence for anyone building on open weights. Whatever framework eventually appears will be written around the closed labs that showed up to the table. Open-weight models will be governed, if at all, by procurement rules and contract terms, not by pre-release review. It is the same conclusion I reached in the frontier model router for 2026: the policy layer is going to treat closed and open models differently, so your architecture should be able to swap between them without a rewrite.

A missed federal deadline does not create a regulatory vacuum. It confirms that the governance you actually operate under is the one you wrote into your own contracts and prompts.

What you actually have to do now

Governance today is contractual and lab-internal. The practical checklist looks like this:

• Read your vendor terms, not the executive order. The usage policies and enterprise agreements from Anthropic, OpenAI and Google are the binding documents. TRAINS participation is voluntary and changes nothing in your contract.
• Document model provenance. For every model in production, record vendor, version, whether the weights are open, and where they were fetched from. This is the field any future review or procurement rule will ask for first.
• Keep a fallback across the closed/open line. If a pre-release review ever delays a closed model you depend on, you want a tested open-weight alternative, and the reverse if a procurement rule ever touches the open one.
• Run your own evaluations. There is no federal benchmark to point to, classified or otherwise. The red-team practice in LLM security red teaming and the output checks in guardrails and output validation are the evidence you will actually be asked for.
• Version your prompts and policies. When an auditor eventually asks what the system did on a given date, the answer lives in your prompt history. Prompt versioning for a small team is the minimum.
• Write the one-page answer for legal now. 'No federal obligation attached to August 1; TRAINS is voluntary; Meta is outside it; here is what we do instead.' Your compliance team will ask within 48 hours, and a prepared answer is worth more than a perfect one.

The honest limitation

'Nothing published by August 2' is a snapshot from Yahoo Finance and TechTimes coverage of the deadline, not a statement about what agencies will do next month. A Federal Register notice could land any day, and the voluntary disclosure framework in particular could appear as a lab-signed document rather than a government one. What is safe to say is that the deadline produced no deliverables, that the disclosure regime for closed models will be voluntary when it arrives, and that open weights sit outside it by design. Plan for that shape, and do not wait for the date to matter.