Anthropic's September 18 evaluator announcement adds a substantial commercial relationship to this investigation. Faculty, now part of Accenture, will lead embedded evaluation of Anthropic models. The work includes alignment assessment, model testing and safeguards. Anthropic will pay Accenture directly. Its discussions with METR about self-funded nonprofit pilots describe a separate arrangement. Anthropic's announcement.
Accenture was already helping expand Claude's market. In December 2025 the companies announced an Accenture Anthropic Business Group: roughly 30,000 professionals to train, joint industry offerings and a co-invested Claude Center of Excellence. That makes commercial expansion part of the assessor's existing relationship with its subject. The earlier partnership.
There is an actual product connecting them. Accenture launched Cyber.AI in March with Claude as its central reasoning engine and reported deploying it in its own infrastructure. Its Agent Shield component addresses oversight of autonomous agents. These are corporate product and deployment statements; the performance figures have not been independently verified here. Cyber.AI.
Faculty brings a specialist business into that arrangement. Companies House records Accenture plc control of Faculty Science Limited from March 12: at least 75% of shares and votes, plus director appointment/removal rights. Completion was announced March 16, with Faculty CEO Marc Warner additionally becoming Accenture CTO. Filed control; completion.
The partners each expect at least $1 billion in AI-safety investment over five years. That is an announced intention, not a $2 billion evaluation fee. Their releases do not supply the publication, removal, dispute or conflict-separation provisions that will govern an adverse assessment. Accenture's announcement.
Anthropic's June policy framework offers a revealing comparison. It proposes evaluator qualification, conflict disclosures, pooled financing and measures against shopping for accommodating reviewers. Those proposals are not the subsequent commercial agreement. The next useful document is that agreement, followed by evidence of how it handles an actual disagreement. Framework, pages 7-8.
METR's separate OpenAI incident review illustrates another dependence. It charged no assessment fee, but estimates using about $400,000 in free OpenAI API credits. METR also says preserving future voluntary access influenced drafting, editing and redaction judgments. It stands by its substantive conclusions. Report, setup and footnotes 5 and 7.
This is where access and obligation matter more than the word independent. An implementation company may bring valuable expertise to scrutiny. Its ability to publish unwelcome findings, retain access and resist commercial pressure determines how much assurance the public receives. Explore the relationships.