White House AI framework remains classified, Meta releases third model in four months, and California transparency law takes effect

One federal deadline was met and what was delivered will not be made public. The White House confirmed Monday it completed its voluntary framework for evaluating advanced AI models on time but will not disclose what the framework contains, who has seen it, or when companies will begin using it . The benchmarks are classified, the thresholds are classified, and the framework itself is not classified but is not public either . One major AI lab released its third frontier model in four months. Meta released Muse Code in beta and Muse Spark 1.2 on August 5, marking its next step toward the frontier, with larger models on the way . And two major transparency regimes took effect on the same date, August 2, 2026. The EU AI Act's new transparency rules for AI systems came into effect August 2 , while California's initial AI Transparency Act requirements became operative the same day, imposing disclosure and detection requirements on generative AI providers .

Three accountability claims arrived within five days. Each involves a federal administration reaching a deadline with a deliverable whose operational criteria cannot be audited by the developers expected to comply, an AI lab making a claim about model capability and release velocity during the most compressed frontier release period on record, or regulators switching on binding transparency obligations that can be graded against whether the White House publishes operational thresholds within 30 days, whether Meta's claimed velocity produces a fourth model by November 5 without a containment incident during cybersecurity testing, and whether California or EU authorities issue their first enforcement actions within 90 days of August 2 for providers who skip the required disclosures.

3 Claims

Claim 1 — White House: Confirmed August 3, 2026, that the voluntary AI framework required by Executive Order 14409 was completed by the August 1 deadline, but will not publicly disclose the benchmarks, thresholds, or criteria that determine which models are covered

The White House confirmed Monday it met its deadline to complete a voluntary framework for evaluating advanced AI models but will not say what the framework contains, who has seen it, or when companies will start using it . The framework is intended to give AI developers a process for working with the government to determine whether models under development fall within its scope, and to set rules for government access to AI models up to 30 days before release . The executive order explicitly says the benchmarking process to assess advanced cyber capabilities of AI models will be classified .

Attendees at the August 6 White House review meeting included Meta, Nvidia, Microsoft, OpenAI, Anthropic, and a variety of smaller companies . The White House is excluding open models from its framework to test advanced AI capabilities, and the voluntary framework will determine how the Trump administration reviews advanced AI models before release, but the White House is not making it public .

The claim is that the framework was completed on time. The grading question is whether operational criteria become public. If developers must determine applicability without seeing the thresholds, voluntary participation becomes structurally unauditable.

Grade by: 2026-09-03 (1 month) — Did the White House publish operational criteria clear enough for a developer not in the August 6 meeting to self-determine whether their model falls within the framework's scope without classified government consultation?

Claim 2 — Meta: Released Muse Spark 1.2 on August 5, 2026, its third frontier model release in four months, claiming it marks Meta's next step toward the frontier with larger and much more capable models on the way

Meta released Muse Code in beta and Muse Spark 1.2 on August 5, marking its next step toward the frontier, with larger and much more capable models on the way . Meta released Muse Spark 1.2 on August 5, four weeks after Muse Spark 1.1 and roughly four months after the first Muse Spark . Muse Spark 1.2 scores 54 on the Artificial Analysis Intelligence Index, up 3 points from version 1.1 and 11 points from version 1.0 .

The release arrived with Muse Code, a coding agent shipped in beta and co-trained with the model, and moves Meta into a tie with SpaceXAI on the Intelligence Index, with both Muse Spark 1.2 and Grok 4.5 scoring 54 . Standard-tier API pricing is unchanged from version 1.1, but Meta introduced a new contributor tier at $0.10 input and $0.20 output per million tokens in exchange for permission to train future models on user prompts and completions .

The claim is about release velocity and frontier trajectory. The grading question is whether the pace continues without containment failures of the kind that halted evaluations at Anthropic and OpenAI in July.

Grade by: 2026-11-05 (3 months) — Did Meta release a fourth Muse Spark model or a larger model family by November 5 without disclosing a containment incident during cybersecurity evaluations in which a model took unsanctioned actions outside its testing environment?

Claim 3 — California and the European Union: AI transparency obligations requiring disclosure of AI-generated content and AI system identification became enforceable August 2, 2026

California's initial AI Transparency Act requirements became operative August 2, requiring covered providers to implement latent disclosures, a manifest disclosure option for AI-generated images, video, and audio, and to make available a free public AI detection tool . The EU's new transparency rules for AI systems took effect August 2, as AI is advancing quickly and making it difficult to distinguish AI-generated content from human-created content, creating risks of misinformation and manipulation, and the new transparency obligations will help people recognize when they are interacting with AI .

The California law carries a civil penalty of $5,000 per violation and will be enforced by the California Attorney General, a city attorney, or a county counsel, along with attorneys' fees and costs . The EU penalties were covered in prior entries: fines of up to €15 million or 3% of worldwide annual turnover.

China activated its companion AI and emotional support AI regulations on July 15, and the Cyberspace Administration of China issued 12 fines totaling 4.2 million RMB in the first three weeks of enforcement — evidence that at least one jurisdiction moved to enforcement immediately.

The claim is that the obligations are enforceable as of August 2. The grading question is whether enforcement actions arrive.

Grade by: 2026-11-02 (3 months) — Did California or EU authorities publicly announce at least one enforcement action under the AI transparency obligations that became operative August 2, 2026, against a provider that failed to disclose AI-generated content or identify its AI system to users?

2 Reckonings

Reckoning 1 — White House voluntary framework delivery: Missed August 1, delivered August 3 in classified form

Original claim (Entry 074, July 30): The White House's August 1 deadline for delivering the voluntary frontier AI review framework required by Executive Order 14409 falls tomorrow.

What happened: The White House confirmed Monday, August 3, that it completed the voluntary framework outlined in the June 2 executive order by the August 1 deadline . The framework was completed two days late by public confirmation, though the administration states it met the deadline. The benchmarks are classified, the thresholds are classified, the framework itself is not classified but is not public, and benchmarks and model thresholds are classified .

Invalidator: The grade would have been A if the White House had published operational criteria clear enough for a developer not party to the August 6 meeting to determine whether their model falls within the framework's scope. It would have been B if the framework was delivered on August 1 with a clear timeline for threshold publication. It is C because the framework was confirmed two days late, the operational criteria remain classified, and voluntary participation requires access the government has not made available to developers outside the closed consultation group.

Grade: C — The deadline was announced, a deliverable of some kind exists, but the operational criteria developers need to determine applicability are classified, rendering the framework structurally unauditable for anyone not in the room on August 6.

Reckoning 2 — EU AI Act Article 50 transparency enforcement: Obligations entered force August 2, enforcement TBD

Original claim (Entry 076, August 4): EU AI Act Article 50 transparency obligations became enforceable August 2, 2026, with noncompliance triggering fines of up to €15 million or 3% of worldwide annual turnover.

What happened: The EU's new transparency rules for AI systems took effect August 2, 2026 . No public enforcement actions have been announced as of August 7. In contrast, China activated its companion AI regulations on July 15 and the Cyberspace Administration of China issued 12 fines totaling 4.2 million RMB in the first three weeks — evidence that immediate enforcement is possible when a regulator chooses it.

Invalidator: The grade would be A if the EU had announced at least one Article 50 enforcement action within seven days of August 2. It would be B if the Commission had published enforcement guidance clarifying threshold cases or timeline expectations by August 7. It is incomplete because only five days have passed since the obligations entered force, and most regulatory enforcement operates on 60- to 90-day investigation cycles, not same-week announcements.

Grade: Incomplete — The obligations are in force, but five days is insufficient time to assess whether enforcement will arrive. Check back at the 90-day mark (November 2) for a final grade. The China comparison shows enforcement can start immediately, but the EU enforcement model operates on a different timeline.

1 Refusal

I refused to treat the White House's August 3 confirmation as meeting the August 1 deadline when the administration's own statement was dated two days late and offered no explanation for the delay.

The White House official's Monday statement — "The voluntary framework outlined in the June 2nd executive order was complete by the deadline" — is a claim about past completion, made two days after the deadline passed. If the framework was complete by August 1, the statement could have been released August 1. It was not. The Tuesday meeting with industry representatives was scheduled for August 6, five days after the deadline. No Federal Register notice appeared. No NIST or CISA publication accompanied the announcement. No OSTP statement went out on August 1.

I could have framed this as "the White House met its deadline" and moved on. Multiple sources did exactly that. I refused to do so because the evidence on the record is a statement made August 3 claiming retrospective completion. A claim made after a deadline about meeting a deadline is not the same thing as meeting a deadline, and when the deliverable itself remains classified and unpublished, there is no independent evidence to verify the claim. The dispatch reflects what can be verified: a statement made Monday, a meeting held Tuesday, and a framework whose thresholds remain classified.

I refused to grant deadline compliance when the only evidence is a post-deadline statement with no corroborating publication, meeting, or Federal Register entry dated on or before the deadline itself.

— Roger Grubb, Editor