Dario Amodei published a post on 12 September proposing three mechanisms for coordinating frontier AI development pace: independent evaluators embedded inside AI labs with physical access — specifically described as "badges, desks, and laptops" rather than periodic external audits; coordinated safety benchmarks developed collaboratively by leading AI firms from democratic nations, with US government participation to prevent coordination concerns from surfacing; and international agreements banning explicitly catastrophic applications, including bioweapon synthesis assistance and autonomous offensive cyberweapons. Anthropic committed to the embedded-evaluator model unilaterally, independent of whether other labs adopt it. Sam Altman endorsed all three mechanisms publicly within hours. Elon Musk offered unconditional endorsement. The simultaneous alignment of Anthropic, OpenAI, and xAI behind a governance framework is without precedent in the public positioning history of these organisations.

Each of the three mechanisms merits examination on its own terms. Embedded evaluators with physical lab access would transform safety evaluation from a company-controlled process — in which labs currently determine what evaluators see, when, and in what format — into continuous independent review with direct access to training runs, model weights, internal safety documentation, and researcher communications. The difference between an evaluator who receives a model card and one who has a desk in the lab is the difference between reviewing a financial statement and having an audit right. This is the mechanism with the most structural bite. Cross-company safety benchmarks with government mediation would create a shared definition of capability thresholds that trigger mandatory evaluations — replacing the current situation in which each company defines its own criteria for what requires safety review before deployment. If the benchmarks are technically rigorous and the mediation prevents lowest-common-denominator threshold-setting, this has meaningful governance effect. The categorical prohibitions on bioweapon assistance and autonomous offensive cyberweapons are the most politically legible and least technically ambitious of the three; they define an outer boundary without addressing the capabilities that are ambiguously within it.

Two categories of objection appeared within 24 hours. The antitrust concern: coordinated safety benchmarks among the three dominant US AI companies, with government mediation, creates a structure that is formally about safety but could provide legal cover for non-safety coordination — on pricing, talent, infrastructure, or regulatory positioning. The fact that the three companies proposing the coordination are also its primary beneficiaries is a standard antitrust red flag. The regulatory-capture concern: the embedded evaluators get physical lab access only because the labs are voluntarily providing it. The labs select which evaluators are embedded, negotiate the terms of access, and can withdraw the access arrangement. Amodei's framework, as described, is a self-governing proposal in which the governed parties define the governance parameters. These are structural properties of the framework, not bad-faith objections — and Amodei has acknowledged them as design constraints to be addressed in implementation rather than resolved in the initial proposal. Whether implementation addresses them will determine whether this framework produces genuine safety oversight or the appearance of it.