Nation Runner

The badge on the desk

Dario Amodei’s pacing essay does not ask the public to trust Anthropic. It asks the public to trust a third party with a badge, a laptop, and a view of the training run. The name he uses is METR.

That is the rabbit hole. Not “is safety fake.” Not “are evals useless.” Who pays the people who would sit in the lab if “embedded evaluators” became the law of the frontier.

METR says, plainly, that it does not take cash from AI companies, and that it will not take donations made by or at the direction of their staff. That sentence is doing a lot of work. It is also not the whole map.

Source documents

METR’s own About page (retrieved 15 September 2026) is the primary ledger for this file:

METR is funded by donations. We are grateful to METR’s many supporters: from METR’s first institutional-scale funding through The Audacious Project (a funding initiative housed at TED); to individuals from Jane Street; foundations such as the Sijbrandij Foundation, The Pew Charitable Trusts, Schmidt Sciences, the Packard Foundation, LaCentra-Sumerlin Foundation, Astralis Foundation and Expa.org; the AI Security Institute; pooled funds such as those of Longview Philanthropy and Effektiv Spenden; recommendations by the Survival and Flourishing Fund; and many others, such as David Farhi, Geoff Ralston, Dylan Field and Steve Newman.
(Note: METR cannot accept donations made by or at the direction of frontier AI company employees.)
Additionally, a small part of our income is from a technical assistance contract with the European AI Office… METR has not accepted funding from AI companies, though we make use of significant free tokens, which we use for evaluations, research, and engineering.

On 14 August 2026 METR published a funding update:

In the last 6 months, METR raised commitments of around $71 million.

Same post, same independence line, then the residual:

However, frontier AI companies currently provide a significant amount of free tokens for our evaluations, research, and engineering.

Partnerships listed on the same About page: OpenAI, Anthropic, Google DeepMind, Meta, and Amazon — “to pilot frontier risk assessments.” Those companies “have also provided access and tokens.” METR is in the NIST AI Safety Institute Consortium and the California Cybersecurity Task Force, partners with the (UK) AI Security Institute, and assists the European AI Office.

This file does not invent a secret cheque from Anthropic. It maps the public graph Amodei’s plan would deputize.

ARC spawn, not a random NGO

METR is not a stray consumer-protection shop that wandered into Palo Alto. It is ARC Evals with a new name.

On 4 December 2023 the team announced that ARC Evals was leaving incubation at the Alignment Research Center and becoming Model Evaluation & Threat Research — METR, “pronounced meter,” a nod to metrology. Beth Barnes stayed in charge. Paul Christiano, who founded ARC, declined a board seat because he was heading safety at the US AI Safety Institute.

That lineage matters for the pacing plan. ARC was a theory shop. ARC Evals was the empirical arm: measure what frontier models can actually do, especially long-horizon agentic tasks. METR’s public mission is still catastrophic-risk measurement — autonomous capabilities, self-improvement, shutdown resistance — and it prototyped Responsible Scaling Policies, which it says nine developers have adopted.

The priesthood Amodei wants “embedded” is not “the public.” It is this professional class: people who already believe the risk is civilizational, who already write the tests that would tell a regulator when to slow a run, and who already sit in the same safety conversation as the labs they would badge.

Ajeya Cotra is listed on METR’s About page as technical staff. That is not a smear. It is a graph. Open Philanthropy’s worldview — long timelines, compute-based forecasting, pause-adjacent philanthropy — is in the building. Yoshua Bengio is listed as an advisor. The UK AISI is a partner. The European AI Office is a client. If you need an evaluator the state can point at, this is what the Rolodex looks like.

Follow the disclosed money

Start with what METR published, not what a critic wishes were true.

Philanthropy, not product revenue. METR is a 501(c)(3). The About page does not sell evals as a business. The cash is donations and a small EU technical-assistance contract.

First institutional-scale cheque: The Audacious Project. That is TED’s big-bet vehicle. METR names it as the start of institutional-scale funding. Later secondary write-ups put a large Canary / RAND-adjacent allocation in 2024; this file sticks to METR’s own words for the dollar claim that is dated and first-party: about $71 million in commitments in the six months before 14 August 2026.

The rest of the named board. Jane Street individuals. Sijbrandij (GitLab money). Pew. Schmidt Sciences. Packard. LaCentra-Sumerlin. Astralis. Expa. Longview. Effektiv Spenden. Survival and Flourishing Fund recommendations. Named people: David Farhi, Geoff Ralston, Dylan Field, Steve Newman. UK AISI appears in the supporter list as well as the partner list.

Read that list as a coalition, not as a conspiracy. These are funders who already think frontier AI is a civilisation-scale measurement problem. They are not GM or Exxon writing a trade-association cheque. They are the pause-adjacent donor class: effective-altruist pooled funds, longtermist vehicles, a handful of tech-fortune foundations, a state safety institute, and a TED project that likes “independent science” as a headline.

If embedded evals become a licensing layer, this is the endowment behind the license.

What is missing, on purpose. OpenAI, Anthropic, Google, Meta, Amazon: not on the cash list. METR repeats that as policy. Treat it as true until a filing says otherwise. Then notice what replaced the cash.

The token residual

Free tokens are not a rounding error in an eval shop. They are the feedstock.

METR’s product is measurement of other people’s models. Those models live behind lab APIs. “Significant” free tokens from the frontier labs means METR’s cost base is subsidised by the firms it would sit inside. Access is also listed: not just a public API key, the kind of access you need to run pre-deployment assessments.

Amodei’s Step 1 is the same access, made permanent and physical: desks, badges, training pipelines. METR already has a version of the relationship. The pacing essay would turn a research partnership into a required occupancy.

You can believe METR is sincere and still see the structure. A lab that wants a slowdown with a verifiable stamp has a preferred stamp-issuer. The stamp-issuer will not take the lab’s grant. It will take the lab’s compute, the lab’s weights-under-NDA, and the state’s blessing. That is a cleaner story than “Anthropic paid METR.” It is also more durable. Compute and access are how you capture an evaluator without appearing on a Form 990 as a donor.

METR’s own conflict-of-interest page exists because they know this. Policy is not a solvent. The residual remains: the only people who can give you the thing you measure are the people you are measuring.

Who benefits if the badge becomes law

Return to the sibling file: Anthropic’s pacing plan. Step 1 is unilateral and cheap for a frontier lab that already hosts evaluators. Step 2 is industry coordination with a possible antitrust waiver. Step 3 is global, optional, and the part that would actually bind China.

If Step 1 is written into US or EU rules, METR-shaped orgs become a licensing layer. Licensing layers need funders who can scale headcount. $71 million in half a year is a scaling round dressed as philanthropy. It hires the recursive-self-improvement team, the monitoring team, the incident team — the exact functions a regulator would want on site.

Pause funders win a standard. The donor class listed above already paid for the measurement regime. Making that regime mandatory socialises their theory of risk. Rival theories (open-weight, accelerate, “the risk is concentration not FOOM”) do not get a desk on the training floor.

Incumbent labs win a costly signal they can already afford. Anthropic can host METR. A garage lab cannot. Free tokens flow to the evaluator from the same five names on the partnership line. Challengers still have to buy the GPU hours METR gets comped.

The state wins a private bureaucracy. UK AISI, US NIST consortium, California task force, EU AI Office contract: METR is already a contractor-shaped object. Embedded evals are how that object enters the lab without the government hiring 400 people onto a GS scale.

METR wins permanence. Philanthropic commitments are lumpy. A legal duty to have “independent” people in the building is an annuity. The About page’s independence sentence becomes the brand the annuity requires.

None of that requires a secret meeting. It is what you get when the only politically available “independent evaluator” is a nonprofit whose donors, advisors, alumni, and compute suppliers are the pause coalition plus the labs being paced.

What this file does not claim

It does not claim METR falsifies evals. It does not claim the $71 million is laundered lab money. It does not claim Barnes is an Anthropic employee. METR’s published rule against lab cash and lab-directed donations is part of the evidence, not something to skip.

It claims this: “independent” in the pacing essay is a sociological fact, not a thermodynamic one. The cash is philanthropic and pause-adjacent. The compute is free from the frontier. The institutional parents are ARC and a safety-institute Rolodex. The named customer in Amodei’s Step 1 is this shop. If you are going to pace the frontier with badges, you should be able to draw that map without flinching.

The honest residual is the same as in the pacing critique. Some outside measurement of autonomous capabilities is reasonable. The danger is converting that measurement into a gate staffed by a single professional class, funded by people who already want the gate, hosted by the firms that would be gated, and then asking Washington to make defection illegal.

Follow the money, including the money that arrives as tokens. Then ask who gets the badge.