Anthropic’s Framework governs the most dangerous AI. It is not governing the most common AI use case.
Rita Felgate | August 2026
ruleoflaw.science
An em-dash publication: developed through human-AI cognitive synthesis with Claude (Anthropic). Learn about the em-dash methodology
Something is missing from every serious AI governance framework currently in circulation. Not from malice. Not from oversight. The frameworks are addressing the risks they were designed to address. But the most common AI use case — AI drafting, processing, and operationalising the instruments of state power — is not in any of them.
That is the hole in the bucket.
The Framework and the Use Case It Does Not Address
Anthropic’s Advanced AI Framework, published June 2026, is serious work. It addresses catastrophic risks from frontier AI — biological weapons, offensive cyber operations, loss of control of AI systems, and automated research and development. It proposes mandatory independent evaluation, transparency obligations, and enforcement authority. It is the kind of framework that governance of dangerous AI requires.
It is entirely silent on the most common AI use case.
The AI that will affect most people most often is not the frontier model designing bioweapons. It is the AI system drafting the regulation that sets the fine for your expired licence, processing the ministerial notice that determines your building permit, generating the commencement notice that brings an Act into force — or fails to. Blog 2 in this series showed what the Rule of law requires of AI when it enters that chain. Every one of those instruments must satisfy the Seven Rule of law (ROL) Compliance Categories — Authority, Jurisdiction, Clarity, Public Participation, Publication, Referent, Commencement. These conditions are necessary to the coordination capacity of the state and hence its stability. Anthropic’s Framework addresses none of this.
This is not a criticism of Anthropic. It is an identification of a different problem — one the Framework was not designed to solve and cannot solve without a measurement standard it does not currently have.
Independent Evaluation Needs a Standard
The Framework proposes mandatory independent evaluation of AI systems. That is the right instinct. But the Framework scopes its evaluation to four Enumerated Risks — biological weapons, offensive cyber operations, loss of control of AI systems, and automated research and development. None of those is the documentary compliance of governance instruments. The Framework itself says it wants evaluator standards published. For the use case this blog addresses, that standard does not yet exist.
Consider the structure of any governance preservation infrastructure. It records which standard was applied and what result followed. What it cannot supply is the standard itself — the observable conditions for ROL-compliant coordination that any governance instrument must satisfy. Preservation without an upstream standard produces a record of whatever was submitted, compliant or not. The standard must come from upstream.
Without an upstream standard, third-party evaluation of AI governance instruments produces accountability infrastructure without accountability content. The evaluation is preserved. What it evaluates against remains undefined.
The Seven ROL Compliance Categories are that upstream standard — published open access, empirically grounded against 28,337 gazette instruments (Felgate, R. (2026). DOI: 10.5281/zenodo.21134975), available today. A third-party evaluator applying the Seven Categories to AI-generated governance instruments produces a verifiable finding: pass or fail on each observable condition, overall verdict, identified coordination defects. That is not another opinion. That is a measurement.
Government Contract Termination Without Documentary Specification Is Not Governance
Several European governments have restricted or ended contracts with AI suppliers in governance functions — France, Germany, Spain among them. The stated reasons are supplier-relationship risk: dependence on partners capable of cutting access to critical technology at will. Those states can exit a supplier on those grounds. What they cannot do is state what documentary compliance standard the next supplier must meet. That is the governance gap. A state can name the relationship risk. It cannot yet name the Category failure.
Terminating a contract without specifying which documentary conditions the supplier failed to satisfy is reaction, not governance. ROL-compliant governance depends on prospective specification of the conditions that make AI’s exercise of state power legitimate — not retrospective termination when something feels wrong. A state that cannot name the Category failure cannot govern the supplier.
The ROL due diligence clause — introduced in Blog 2 in this series — gives governments the prospective specification they need. It requires AI suppliers to verify and document Seven Category compliance at the point of production. It makes the documentary basis of AI’s exercise of state power visible before the contract ends, not after.
Governments are currently at the stage of recognising the problem. The measurement standard that lets them name it precisely already exists.
The First Step Anthropic Can Take Today
In June 2026, California made Claude the first AI tool available to all its state agencies and local governments at half price. The Governor’s office stated explicitly that Claude will help state workers draft documents and analyse information. That is the SROL’s exact use case, named by the customer, in a live Anthropic contract, at a scale where a pilot clause is actually negotiable.
The ROL due diligence clause exists the moment two parties agree to it. It does not require new legislation. It does not require a new regulatory body. It requires an AI supplier willing to warrant the production-side Categories — Clarity, Publication, Referent — and a government willing to warrant the process-side Categories — Authority, Jurisdiction, Commencement — with both parties verifying against the documentary record that Public Participation occurred. The allocation of Categories between supplier and state is developed in Blog 2 in this series.
California is drafting documents with Claude today. The clause that makes that activity verifiably ROL-compliant could be in the next contract.
Why This Matters for the Framework Itself
Anthropic’s Framework is designed to prevent catastrophic risks from AI. But those catastrophic risks do not arise in a governance vacuum. They arise in governance systems whose information architecture is already compromised — where the instruments of state power systematically fail the Seven ROL Compliance Categories, where coordination capacity is degraded, where the state’s ability to detect and respond to emerging risks is already weakened by years of accumulated Category failures.
A state whose governance instruments fail the Seven Categories is a state whose incident reporting mandates, independent evaluation requirements, and enforcement mechanisms are themselves built on a defective documentary foundation. The Framework’s catastrophic risk governance depends on governance instruments that must themselves be ROL-compliant to function. Securing the ordinary end of AI governance is the foundation that makes the catastrophic end governable.
The Framework governs the most dangerous AI. One clause governs the most common AI. Both are necessary. One already exists.
The Seven ROL Compliance Categories are published open access and available for any AI supplier or government to apply today. If the standard your framework is missing is the one described here — get in touch.
The Science-based Rule of law framework has been in development for twenty years. Its systematic documentary measurement covers sixteen years of South African government gazette data, applying formal scientific methodology to governance. The Seven ROL Compliance Category methodology is published open access: Felgate, R. (2026). DOI: 10.5281/zenodo.21134975.
Rita V. Felgate is an independent legal practitioner and governance researcher. She is the founder of ruleoflaw.science and the developer of the Science-based Rule of law framework.
Leave a Reply