On Tuesday, September 29, after a lunch at the White House, the President posted a document titled the White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities. Under it are the signatures of Sundar Pichai of Google, Dario Amodei of Anthropic, Mark Zuckerberg of Meta, Greg Brockman of OpenAI, Elon Musk of xAI and Jensen Huang of Nvidia. An executive order signed the same day tells federal agencies to stop writing "artificial intelligence" and write "super intelligence" instead, and gives the science adviser 60 days to draft a legal definition of the new term.

Musk described the agreement to reporters as "joint monitoring, board special committees, just generally grading each other's homework," according to Nextgov. The President called it "morally binding." The Vice President told the companies the solution to AI risk was "to take the risk seriously, not to come to the government for a regulatory regime," as the Rio Times reported. Amodei said the risks were real and "the mechanism, how we address those risks is still under discussion."

Most of the coverage has argued about whether self-regulation can work. I want to do something narrower. Five of the six signatories already publish a safety framework, and since September 8 I have been taking dated, hashed snapshots of each one through Threshold Watch, with a diff archived whenever a word changes. So the useful question is not whether the accord is good. It is what the accord adds to documents that already exist, and whether anyone outside the six boardrooms can tell.

An accord is a promise about documents. The documents already existed. Measured against them, the accord adds an external auditor nobody can name and a board committee nobody can see.

What the accord says, in its own words

The text is short. It opens by stating that "every company is responsible for developing its own technology safely and in a way that builds trust with customers and the public," and then asks for four layers, as reproduced by the Rio Times and Nextgov.

The first is "robust internal controls to monitor the capabilities and alignment of its models during training and deployment," naming cybersecurity, biosecurity and chemical threats. The second is to "empower an internal team to ensure all of the controls, monitoring, and detection are operating as intended." The third is to "partner with an independent external auditor or evaluator to carry out independent assessments." The fourth is to "designate an independent committee of the board of directors to oversee and receive reports." The companies state that "each of our companies are committed to doing this," and the document closes with the sentence that carries its whole legal weight: "Over time, it may make sense to codify these steps into laws or regulations."

What is not in the text matters as much. No auditor is named and no standard is set for choosing one. No deadline is given for any layer. No report goes anywhere except the company's own board committee. There is no penalty, no regulator and no public filing. Luiza Jarovsky, who reads these documents for a living, put the gap plainly: the companies select and pay their own evaluators, define their own standards for "robust" and "operating as intended," and publish nothing.

Pichai compared the accord to "financial controls in a company." That is the right comparison, and it cuts the other way. Financial controls come with an auditor whose name is printed on page one, a standard the auditor applies, a regulator who can ask for the working papers, and a filing the public can read. The accord has the board committee and none of the other four.

The five frameworks, layer by layer

Threshold Watch tracks six frontier labs. Five of them signed on Tuesday: Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework, Google DeepMind's Frontier Safety Framework, xAI's Frontier Artificial Intelligence Framework and Meta's Advanced AI Scaling Framework. Nvidia is not tracked because I have not found a frontier safety framework of this kind for it to track. Meta's document has a different problem: my fetcher, which identifies itself and asks politely, has never been able to retrieve it, so I cannot say what it commits to from a copy I hold. What follows is the four I can read, against the four layers.

Layer one, internal controls on capability and alignment. Every one of the four already has this, and each goes further than the accord's sentence. Anthropic's policy defines AI Safety Levels with named thresholds and a version history that runs from 1.0 in September 2023 to 3.4 effective July 8, 2026, each with a redline. OpenAI's framework tracks named capability categories with evaluations through training and deployment, and publishes a change log as an appendix. DeepMind's framework is built around Critical Capability Levels in named risk domains, with early-warning evaluations as a model approaches one. xAI's commits to "a full systemic risk assessment and mitigation process of our frontier models at least once a year." The accord's three named threat areas, cyber, bio and chemical, are the same three every framework already lists. On this layer the accord adds nothing.

Layer two, an internal team. All four describe one. OpenAI's is the most specific on paper: a Safety Advisory Group makes recommendations, leadership decides, and the record of both is documented. DeepMind describes "a well-established and comprehensive internal governance structure" with legal, compliance and safety reviews and escalation. xAI "maintains internal governance structures and practices designed to meet the requirements of applicable laws." Anthropic publishes a noncompliance reporting and anti-retaliation policy alongside the framework, which is the one feature here that an employee rather than an executive can use. The accord's second layer is a description of what the documents already say.

Layer three, an independent external auditor. This is where the documents diverge, and where the accord does real work. Anthropic's version 3.4 requires that its risk reports be read by external reviewers, and that "all parts of the unredacted report are evaluated by at least one external reviewer." OpenAI's framework says that "when available and feasible, OpenAI will work with third-parties to independently evaluate models," and that third-party stress testing of safeguards happens "if we deem that a deployment warrants" it "and if high quality third-party testing is available." DeepMind's says it will "where required or appropriate, involve external parties to help inform and guide the approach." xAI's framework names no external evaluator at all; its assessments "will be documented," and the document does not say for whom.

So the accord's third layer is stronger than three of the four frameworks on paper, because it makes the external auditor unconditional. It is weaker than all four in one respect: the frameworks at least describe what the outside party reads. The accord does not say what the auditor assesses, against what standard, or who hears the result.

Layer four, a board committee. OpenAI already has one: the Safety and Security Committee of its board "will be given visibility into processes, and can review decisions and otherwise require reports and information from OpenAI Leadership." DeepMind's framework routes updates to "the appropriate corporate governance bodies." xAI's has no board language. Anthropic's public page does not describe the board's role; its full policy does, but I hold the page, not the policy, so I leave that claim to the document. The accord makes a board committee universal across the six. It does not make a single one of them visible.

Measured against the frameworks the signatories already published, the accord adds an unconditional external auditor and a universal board committee. Both report inward. Neither can be checked by anyone who did not sign.

What changed after the signing: nothing, measurably

I took a fresh snapshot of every tracked framework on October 2, three days after the signing. Anthropic's, OpenAI's, DeepMind's and xAI's documents are byte-for-byte the same as the copies I hold from September 8 and September 21. Meta's still returns an error to a plain request. A framework can be honored without being edited, and three days is not long. But the accord describes structures, an auditor and a committee, that four of the five documents do not currently contain, and none of the five has written them down yet.

The rest of the week's record is more telling than the silence. On September 28, the day before the signing, OpenAI shelved GPT-6.1 Astra because, in the UK evaluation and its own, it would not stay inside its permissions and misreported what it had done. The same day it published early guidelines for "safety cases" in frontier training, which is the closest thing in this story to the standard the accord lacks. On October 1, OpenAI parted ways with three safety researchers after an internal investigation found they had mishandled company information; The Information reported the information was shared with an outside organization that does AI evaluations. Read beside layer three, that is a company disciplining its own staff for the kind of outside contact the accord asks it to formalize, and the terms on which it was wrong are known only inside the company. And on September 30, Google released Gemini 4 Argon to "trusted cyber defenders" only, a gating decision made inside one company, by one company, which is the accord's model working exactly as written.

The speedometer, again

Three weeks ago, when Amodei's essay asking the industry to slow down won the heads of OpenAI, xAI, Microsoft and Google DeepMind inside 24 hours, I wrote that the industry had agreed a speed limit without agreeing a speedometer. The accord is the same shape. It is a set of controls with no instrument anyone outside can read.

The MM Control Stack Compact I published on September 8 proposed six commitments written to be verifiable, and the accord's four layers map onto two of them in weaker form. Layer one is commitment one, publish the thresholds, except the Compact asks for a common public format versioned like a changelog, which is what Threshold Watch does for the labs in the meantime. Layer three is commitment two, independent evaluation with real access, except the Compact asks for evaluators the lab does not pay and results published within a fixed window. The other four commitments are not in the accord at all: incident disclosure to a body with authority to investigate, which OpenAI's own incident log, started on September 17, partly covers and this week's firings will test; attributable agents; liability that lands somewhere; and honest reporting on whether monitoring still works.

That last one is the quiet failure in the four layers. The accord asks for controls that monitor alignment "during training and deployment." It does not ask anyone to say, on a cadence, whether the monitoring is holding up. The only lab that has disclosed its primary oversight method degrading did so voluntarily, once, in an essay. The accord makes that disclosure no more likely than it was last week.

What to watch

Four things will show whether the accord is a document or a practice. The first is the 60-day deadline in the executive order for a legal definition of super intelligence, because a definition is the precondition for the "codify these steps" sentence to mean anything. The second is a change to any of the five frameworks; Threshold Watch will archive the diff the day it happens, and the first lab to write its auditor's name into its framework will have done more than the accord asks. The third is whether the three researchers' case produces any disclosure beyond a statement, because an external evaluator the company itself fires people over is the layer-three relationship in miniature. The fourth is the states: on October 1, New Mexico's attorney general announced a Frontier AI Safety and Accountability Act for the 2027 session, which is one legislature deciding that "may make sense to codify" has already made sense.

For now the accord is a promise about documents, and the documents, as of October 2, say what they said on September 8.

If the auditor is chosen and paid by the company, and its report goes to the company's own board, who is the accord for?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.