← All posts
nonprofits, form 990, grants, Coefficient Giving, Tarbell, METR, rematchability

What rematches, what doesn't: the Tarbell–Coefficient grant graph, row by row

Peter Johnston · September 15, 2026

There is a fight underway about AI regulation, and it runs on grant graphs.

On one side: safety advocates, lawmakers, and at least one major lab publicly asking to be regulated. On the other: a growing chorus calling the safety push astroturf — incumbents building a regulatory moat against open-weight models that keep closing the capability gap and keep winning on price. Sincerity and strategy are not mutually exclusive, and from the outside they are nearly indistinguishable. Everyone in this fight, on every side, has a revenue line attached.

That is where this blog gets off. Motive does not appear on any schedule of a Form 990-PF. Neither does coordination. What does appear is money — amounts, dates, payees — and money has one property no thread on X has: it can contradict you. The narrow ground worth holding is the documents. This post holds it.

Kevin Bass — by his own disclosure, a public critic of Anthropic and Coefficient Giving — maintains a GitHub repository of grant-graph figures built to argue that money from the Moskovitz funding orbit reaches METR, the AI-evaluation nonprofit, through indirect channels: its former parent, its project partners, a pooled-fund donor, a board member’s organization, and a journalism program whose fellows have covered METR’s leadership without, by the pack’s account, telling readers about the funding. The pack’s largest number is not a grant at all: a donated Anthropic stake that, on the pack’s own ceiling math, could be worth up to $7.7B — with no public filing identifying where it sits. Two things make the pack worth taking seriously. Every number traces to a named row in a public research CSV, with a source URL and a verbatim quote. And the pack states its own limits: “none found” always travels with the sources and dates it was found in.

This post audits one slice of it — the six Tarbell rows, TB01 through TB06, behind the claim that Coefficient Giving funds the Tarbell Center for AI Journalism. We did not re-derive the pack’s ARC, RAND, or Longview rows, and we assert nothing about anyone’s motives. Motive is parked.

The cast, in one breath. Coefficient Giving is the grantmaker — the organization formerly known as Open Philanthropy, as the redirect of its old Tarbell grant page quietly confirms. Good Ventures Foundation is the private foundation in the same funding orbit whose Forms 990-PF form the public cash trail. The Tarbell Center for AI Journalism is the grantee. The Survival and Flourishing Fund publishes recommendation lists — it recommends, other vehicles pay. METR is Model Evaluation and Threat Research. Players Philanthropy Fund is the fiscal sponsor that keeps appearing on the cash trail.

The short version

The claim that checks out checks out to the dollar. Three Tarbell-named Coefficient awards — $816,000 on 2024-11-17, $2,888,000 on 2025-03-06, and $1,587,930 on 2025-07-15 — sum to $5,291,930, exactly the total in Bass’s package, and nothing in the checked primary documents contradicts any of the six rows. Tarbell’s own website corroborates the relationship from the grantee side.

The two claims that would make the graph explosive — a direct Coefficient or Good Ventures grant to METR, and a paper trail for the donated Anthropic stake — are the ones with no row in any document we checked. On METR, Coefficient’s own published statement agrees with the absence. On the stake, the absence is the finding, and it has a closing date: the next Good Ventures filing.

How a rematch works

A rematch is deliberately dumb. Each row of the claim package is checked against the primary document it cites — no more, no less. A KEEP is a primary document carrying the same amount, date, and grantee. A KILL is a primary document that contradicts the row. Anything that can be neither confirmed nor killed stays OPEN, and the claims are frozen as of 2026-09-15 so the target cannot move while we shoot. The documents: Coefficient Giving’s public grants index (snapshot 2026-09-11, 2,911 rows), Tarbell’s About and Ethics pages, SFF’s 2024 and 2025 recommendation lists, and Good Ventures Foundation’s Forms 990-PF through fiscal year ending June 2025.

The six rows

Table 1 is the whole package, with what the rematch found.

Row What the package claims Rematch In the $5.29M?
TB01 Coefficient → Training for Good, “Operating Costs and Tarbell Fellowship,” $999,000, 2023-09-26 Checks out — joint award to Training for Good No — not a Tarbell-named grant
TB02 Coefficient → Tarbell Center, General Support, $816,000, 2024-11-17 Checks out Yes
TB03 Coefficient → Tarbell Center, General Support, $1,587,930, 2025-07-15 Checks out Yes
TB04 Coefficient → Tarbell Center, Operating Costs, $2,888,000, 2025-03-06 Checks out Yes
TB05 SFF 2024 recommendation, $520,000 via Players Philanthropy Fund Checks out — as a published recommendation, not a paid grant No — SFF, not Coefficient
TB06 SFF 2025 recommendation, $783,000, including a $200,000 conditional match Checks out — as a published recommendation; match fulfillment unverified No — SFF, not Coefficient

Table 1. The Tarbell package (TB01–TB06) against the primary documents. Six checks, six KEEPs, zero KILLs.

The arithmetic is 816,000 + 1,587,930 + 2,888,000 = 5,291,930. Delta versus the package: $0.

Two counting judgments deserve daylight, because a sloppier audit would fumble both. TB01 is real money attached to Tarbell’s name — the fellowship — but the award runs to Training for Good, so it cannot be honestly summed as a Tarbell-named Coefficient grant. And TB05 and TB06 are SFF recommendations: $1,303,000 recommended, not granted by Coefficient, never summed into the total. The $200,000 conditional match inside TB06 is part of the $783,000, not on top of it. The pack’s own rule — never sum money types with each other — is the right rule, and the total obeys it.

The money is real — and corroborated from the other side

The index snapshot is one document; Tarbell’s website is another, with independent reasons to exist, and it points the same direction. Tarbell’s About page lists Coefficient Giving as a $1M+ supporter across 2023–2026. Its Ethics page goes further: as of 2025, the majority of Tarbell’s funding originates with Coefficient Giving.

Say precisely what this establishes. The money is real: the grantmaker’s index and the grantee’s own disclosure carry the same relationship, and the order of magnitude is consistent with the index rows. What it does not establish is anything about coverage. Whether the funding touches what Tarbell fellows write about METR’s leadership is a question a filing cannot answer — that claim lives in the rest of the pack, unaudited here.

The edge the filings don’t have: METR

The graph’s sharpest arrow would run straight from Coefficient or Good Ventures to METR. No such row exists in the checked documents — and that absence is not ours alone: Bass’s own package logs “Direct: none found.” Our pass over the same 2,911-row index found zero METR-named awards, and no grant row on Good Ventures’ Forms 990-PF through FYE June 2025 names METR or Model Evaluation and Threat Research. Coefficient has said so itself, in print: its December 2025 staff suggestions for individual donors states of METR, “This organization is not a Coefficient Giving grantee.”

Hold “none found” to its bounds. It means none in the documents checked, as of their dates. It does not rule out money that moves through a fiscal sponsor or a donor-advised account without the grantee’s name ever appearing on a public schedule — the pack itself flags the SVCF and NPT accounts that pay Coefficient-recommended grants as unattributed. And it says nothing about the pack’s indirect-channel rows, which run through ARC, RAND, and Longview and are outside this audit. What it does say: the direct edge is one the critic and the criticized already agree is not a row.

The stake that isn’t on any filing

The pack’s biggest number is equity. Moskovitz has said he donated an Anthropic stake to what he calls “our foundation”; the pack values it at up to $7.7B — Forbes’s “less than 0.8%” of the company times the $965B Series H post-money, an upper bound by construction, not a valuation. And here the filing trail simply stops. No investment-schedule line on any Good Ventures Form 990-PF through FYE June 2025 names Anthropic. The legal vehicle holding the stake is unidentified on every public document checked.

Two shelves it could be sitting on. The next Good Ventures return — the first covering a period beginning after June 2025 — could show it, along with TB03’s cash. Or it sits in a donor-advised account at National Philanthropic Trust or Silicon Valley Community Foundation, whose filings do not attribute individual accounts or their holdings — in which case no public filing may ever name it, which the pack itself says wherever it touches them. This is the one genuinely unseen holding in the graph, and it stays OPEN, with a date circled on the calendar.

Where the cash trail runs thin

The gaps, translated. The live Coefficient grant pages for TB01–TB04 now return 404; the Wayback Machine has no snapshots of them; the old Open Philanthropy Tarbell page redirects to a focus-area hub rather than a grant. The dollar rematch therefore rests on the index rows and Tarbell’s own disclosures, not on grant prose anyone can still read. And the 990-PF cash trail is partial by construction — filings lag awards:

  • Good Ventures’ FY2023 return shows a Players Philanthropy Fund line of $287,750 with purpose text naming a Tarbell fellow extension — a partial cash trail in TB01’s neighborhood, not the $999,000 award-date figure.
  • FY2024 shows a PPF line of $816,000, purpose “GENERAL SUPPORT” — the same amount as TB02, with purpose text that does not name Tarbell. Suggestive; not definitive, and we do not treat it as a TB02 disbursement.
  • No $2,888,000 paid line for TB04 appears on the FY2024 schedule, and TB03’s award date of 2025-07-15 post-dates the June 2025 fiscal year end entirely.

None of this is a KILL. All of it is the lag between when money is awarded and when it becomes a line on a public return.

What would settle it

Three documents would move this post’s conclusions. Good Ventures’ FY2026 Form 990-PF — the first filing that could show TB03’s cash and any post-June-2025 investment-schedule change — is the big one; until it is public, F3 stays open. Confirmation of cash actually paid on the SFF recommendations, and of TB06’s conditional match, would convert two KEEPs from “published recommendation” to “disbursed.” And nothing on either list settles the question everyone is actually fighting about. The filings cannot tell you whether the current safety push is coordinated, sincere, or both. They can only tell you what moved.

The thing worth taking away

In a fight where every side arrives with a graph, the only neutral ground is the filing. That ground is smaller than the fight — it has nothing to say about motives, and “nothing found” is only as good as the documents searched. But it has one property no thread has: it can say no. This graph’s Tarbell edges held — every dollar in the headline total is a dollar the grantmaker’s index and the grantee’s website both carry. Its two loudest edges — direct METR money, the Anthropic stake — are precisely the ones the public record does not currently support, and the second of them has a date by which it should.

Public-record rematch of named grant-graph rows. Not an evaluation of Tarbell, Coefficient Giving, METR, or Anthropic, and not a forecast of later filings. Corrections welcome if a primary document moves a KEEP or closes F3.

Appendix — frozen claims, sources

These sentences are frozen as of 2026-09-15. A kill is a primary document that contradicts the sentence.

F1 — Tarbell ← Coefficient

As of 2026-09-15, Coefficient Giving’s public grants index (snapshot 2026-09-11) lists three awards naming “Tarbell Center for AI Journalism” totaling $5,291,930 ($816,000 General Support 2024-11-17; $2,888,000 Operating Costs 2025-03-06; $1,587,930 General Support 2025-07-15), and Tarbell’s own About page lists Coefficient Giving as a $1M+ supporter for 2023–2026; Tarbell’s Ethics page states that as of 2025 the majority of its funding originates with Coefficient Giving.

F2 — No direct GVF / Coefficient → METR (through checked filings)

As of the Coefficient grants-index snapshot dated 2026-09-11 and Good Ventures Foundation Forms 990-PF through FYE June 2025, no grant row names “METR” or “Model Evaluation and Threat Research” as grantee (Bass “Direct: none found”; Quill rematch of the same index found zero METR-named awards).

F3 — Anthropic-share vehicle unidentified on public filings (OPEN)

As of GVF Forms 990-PF through FYE June 2025, no investment schedule line names Anthropic; Bass’s $7.7B figure is an upper bound (Forbes “less than 0.8%” × Series H post-money), not a marked GVF holding; the legal vehicle holding any donated Moskovitz Anthropic stake remains unidentified on public filings (open candidates: post-June-2025 GVF FY2026 return; SVCF/NPT donor-advised accounts).

F3 is OPEN. Nothing in this rematch identifies the holder.

Sources

← All posts