evilrobots.lol / tech / evaluator record

Epoch AI

What it was engaged to test, on what terms, what it found in the document's own words, and what happened afterwards. Every line is a record; every record links to the document it was read from.

▣ THE RECORD nonprofit Philanthropy-funded nonprofits STORE 2026-10-23
  • Founded2021
  • HQSan Francisco
  • Registry idEIN 99-4050541
  • TestsBenchmarks and trends (FrontierMath).
  • Money, as the roster puts itOpen Philanthropy; OpenAI funded FrontierMath; Good Ventures $70,000 for FrontierMath improvements, tax year to 30 Jun 2025.
  • Filesdossier · observatory, sliced to it · graph
Labs tested
OpenAI Anthropic
Models
FrontierMath
Span of engagements
23 January 2025 – 7 April 2026
Terms and findings
1 of 2 documents state a term · 0 quote a finding
Money in, US dollars, stated single-recipient rows
$95,600,897 24 in · 0 outfrom Coefficient Giving · European AI Office · Good Ventures Foundation · Jaan Tallinn · METR · Nat Friedman · OpenAI · Schmidt Sciences · Sentinel Bio · Survival and Flourishing Fund · UK AI Security Institute
Seats, by side
evaluator 5 between 1 6 people: Ajeya Cotra, Ege Erdil, Jaime Sevilla, Marius Hobbhahn, Miles Brundage, Tamay Besiroglu
Afterwards
correction 2 staff move 1

Each count is the number of store records that name this evaluator. An engagement is one lab document naming it for one model; an outcome is an event on the record afterwards; the money, seats and moves are the observatory's rows filtered to it. A cell that cannot be filled says not stated, and each table ends with how many of its rows state each field.

▣ SLICE 40 / 40 records
Lab
Model
Engagement
Terms
Outcome
Funder
Person
year, month or day

Filters combine: any of the values ticked within a dimension, all of the dimensions at once. A table that has no column for a dimension you have set says so instead of showing rows that cannot match. The address bar carries the slice.

What it was engaged to test 2 engagements, OpenAI, Anthropic

In date order, each from the lab document that names the evaluator. Under each: the four terms that document states (access, model stage, publication, payment) and the finding in its words. A term the document is silent on reads not stated; the tags beside the terms are read off their words and are what the console filters on.

    • Access window"OpenAI retains ownership of these questions and has access to the problems and solutions, with the exception of a holdout set"
    • Model stagenot stated
    • Publication"Epoch AI is free to conduct and publish evaluations of any models using the FrontierMath problem set commissioned by OpenAI. However, we cannot share the questions and answers with other parties without written permission from OpenAI"
    • Payment"OpenAI commissioned Epoch AI to produce 300 advanced math problems for AI evaluation that form the core of the FrontierMath benchmark"

    Finding: not stated in the document

    Afterwards: correction
  1. 7 April 2026 Anthropic · no model named benchmark co-build
    Document: Anthropic, Claude Mythos Preview system card, Apr 2026 Its capabilities index forked by Anthropic: "we fork from Epoch AI's implementation of this work, the Epoch Capabilities Index (ECI)"
    • Access windownot stated
    • Model stagenot stated
    • Publicationnot stated
    • Paymentnot stated

    Finding: not stated in the document

Dated 2 of 2Model named 1 of 2Terms stated 1 of 2Finding quoted 0 of 2

The terms, as the documents state them

The same engagements as a grid, one row per document and model, so the silences line up: a document that gives an access window but no payment term, or a finding but no publication term, reads that way here.

DocumentModelAccess windowModel stagePublicationPaymentFindingRecord
Epoch AI, the FrontierMath commissionOpenAI · 23 January 2025FrontierMath"OpenAI retains ownership of these questions and has access to the problems and solutions, with the exception of a holdout set"not stated"Epoch AI is free to conduct and publish evaluations of any models using the FrontierMath problem set commissioned by OpenAI. However, we cannot share the questions and answers with other parties without written permission from OpenAI""OpenAI commissioned Epoch AI to produce 300 advanced math problems for AI evaluation that form the core of the FrontierMath benchmark"not statede53c356823
Anthropic, Claude Mythos Preview system card, Apr 2026Anthropic · 7 April 2026no model namednot statednot statednot statednot statednot stated15e43e308c

Access window 1 of 2Model stage 0 of 2Publication 1 of 2Payment 1 of 2Finding, quoted 0 of 2

What happened afterwards 3 on record

Events a party published after an engagement: an incident on a model this evaluator's engagement covered, a correction or added disclosure, a later finding that cut against an earlier one, a lab's statement that it acted or did not act on the finding, or a move from this evaluator to a lab it graded. A row the source attaches to this evaluator by name is marked named; one that reaches it through the model it tested is marked via the model. The kind is what the record states, not why.

KindDeveloperModelWhat the record statesWhenSourceId
correctionnamedOpenAIFrontierMath"On 2026-06-12, we released a major update, addressing errors in 42% of problems"; "We released v2, which corrected 123 problems in Tiers 1-3 and 12 problems in Tier 4. We also removed 5 problems from Tiers 1-3 and 7 problems from Tier 4"Engagement: e53c35682312 June 2026Epoch AI, FrontierMath Tier 4 (v2)f2ba88ddba
staff movenamedGoogle DeepMindno model namedTamay Besiroglu: Epoch AI (co-founder, associate director); Mechanize (co-founder, Apr 2025) -> Google DeepMind (research scientist, by Sep 2026)Sep 2026Epoch AI; Mechanize; his post75f848d85a
correctionnamedOpenAIGPT-6 AstraIts chief executive on the model's Epoch Capabilities Index score: "An update on Astra's ECI score. This is a major step down from the original score we reported (but still within the uncertainty range we expected). It is explained by lower performance on long-horizon software engineering tasks"16 September 2026Jaime Sevilla on X5d33e90812

Dated 3 of 3Model named 2 of 3Tied to an engagement 1 of 3Evaluator named by the record 3 of 3

The money on record 24 rows · $95,600,897 in, stated single-recipient rows

Rows of the observatory's money table that name this evaluator as recipient (in) or as funder (out), each as the record states it: a grant table line, a contract notice, a Form 990 schedule or the party's own disclosure. A figure quoted in prose is marked and never summed.

DirectionFunderRecipientAmount, as recordedWhenRecordSourceId
inCoefficient GivingEpoch AI$25,113,611Eight grants totalling $25,113,611 (Jun 2022 to Jul 2026), including $8,500,000 (Jun 2025)including $8,500,000Jun 2022party's own pageEpoch AI; Coefficient Givingf97c5b81cf
inCoefficient GivingEpoch AI$68,796,286$68,796,286 (Aug 2026), "a $68 million renewal over 3 years""a $68 million renewal over 3 years"Aug 2026party's own pageEpoch AI; Coefficient Giving2f08995fb2
inSurvival and Flourishing Fund; Jaan TallinnEpoch AI$195,000$195,000 (May 2026)May 2026party's own pageEpoch AI; SFF-2026bb9e1db106
inJaan TallinnEpoch AI$600,000$600,000 (Jan 2025)Jan 2025party's own pageEpoch AI; SFF-2026e6b88d3b90
inJaan TallinnEpoch AI$646,000SFF-2026 recommends $646,000 in all (Moskovitz $542,000 of it)SFF- recommends2026party's own pageEpoch AI; SFF-2026aafd049fe8
inSchmidt SciencesEpoch AIno amount publishedNamed donor (Dec 2025), no amountDec 2025party's own pageEpoch AI08a0bf8e3c
inOpenAIEpoch AIno amount publishedCommissioned and funded FrontierMath's 300-problem corenot datedparty's own pageEpoch AI; Sevilla2555c5299b
inOpenAIEpoch AIno amount published"OpenAI purchased access to the verifiers" (Apr 2026)Apr 2026party's own pageEpoch AI; Sevillae85d2d0a4a
inUK AI Security InstituteEpoch AIno amount publishedEpoch lists UK DSIT as a paying partnernot datedparty's own pageMETR; Epoch AI; Manifund; Generality Labsf46738332c
inEuropean AI OfficeEpoch AIno amount publishedEpoch: a paying partner (2026)2026party's own pageMETR; Epoch AI; Transluce; FAR.AIf59b15efd6
inMETREpoch AIno amount publishedA 2025 software-engineering benchmark engagement, funded by METR2025party's own pageEpoch AI7c3cc5bec8
inNat FriedmanEpoch AI; MECHANIZE; Artificial Analysisno amount publishedFunded a pilot of FrontierMathnot datedparty's own pageEpoch AI; Mechanize; Artificial Analysis; AIUC, Series A; Fortune, Oct 8 2025; Meta memo via CNBC16a5ebe3a8
inNat FriedmanEpoch AI; MECHANIZE; Artificial Analysisno amount publishedfirst-named backer of Mechanizenot datedparty's own pageEpoch AI; Mechanize; Artificial Analysis; AIUC, Series A; Fortune, Oct 8 2025; Meta memo via CNBCcf9b872cbf
inNat FriedmanEpoch AI; MECHANIZE; Artificial Analysisno amount publishedfirst-named of the eight backers on Artificial Analysis's About page, no amount statednot datedparty's own pageEpoch AI; Mechanize; Artificial Analysis; AIUC, Series A; Fortune, Oct 8 2025; Meta memo via CNBC5e85e6667c
inNat FriedmanEpoch AI; MECHANIZE; Artificial Analysis$55,000,000named on AIUC's team page among its $55 million investors, and AIUC's Series A post counts "our $15 million seed round from NFDG"investors, and AIUC's Series A post counts "our $15 million seed round from NFDG"not datedparty's own pageEpoch AI; Mechanize; Artificial Analysis; AIUC, Series A; Fortune, Oct 8 2025; Meta memo via CNBC4aef243df2
inNat FriedmanEpoch AI; MECHANIZE; Artificial Analysisno amount publishedFortune names him among the seeders of Situational Awareness LPnot datedparty's own pageEpoch AI; Mechanize; Artificial Analysis; AIUC, Series A; Fortune, Oct 8 2025; Meta memo via CNBC0291b6507f
inNat FriedmanEpoch AI; MECHANIZE; Artificial Analysisno amount publishedsince June 2025 co-lead of Meta Superintelligence Labs for AI products and applied research, per Meta's 30 June 2025 memoJune 2025party's own pageEpoch AI; Mechanize; Artificial Analysis; AIUC, Series A; Fortune, Oct 8 2025; Meta memo via CNBC835d93c2fa
inGood Ventures FoundationEpoch AI$70,000$70,000 (improvements to the FrontierMath benchmark)improvements to the FrontierMath benchmarknot dated990-PFProPublica, Good Ventures FY2024 Form 990-PFafb66f19f8
inSentinel BioEpoch AI$80,000$80,000 (Aug 2024),Aug 2024party's own pageSentinel Bio, Research & Projects; Epoch AI; Sentinel Bio, RAND0ca4edb7b2
inSentinel BioEpoch AI$85,000$85,000 (Sep 2025),Sep 2025party's own pageSentinel Bio, Research & Projects; Epoch AI; Sentinel Bio, RAND4eee235c5d
inSentinel BioEpoch AI$15,000$15,000 (Mar 2026) on Sentinel's list, and Epoch's own page listsMar 2026party's own pageSentinel Bio, Research & Projects; Epoch AI; Sentinel Bio, RANDcb7f8d0445
inSentinel BioEpoch AIrestated$85,000 (Sep 2025)Sep 2025party's own pageSentinel Bio, Research & Projects; Epoch AI; Sentinel Bio, RAND5b2443a2b0
inSentinel BioEpoch AIrestated$80,000 (Sep 2024)Sep 2024party's own pageSentinel Bio, Research & Projects; Epoch AI; Sentinel Bio, RAND17c9992e24
inEuropean AI OfficeMETR; Epoch AI€1,167,484EUR 1,167,484Lot 3, loss of control risk modelling and evaluationDecember 2025contract noticeTED421931534e

Amount stated 11 of 24Dated 16 of 24Purpose stated 6 of 24Funder profiled 24 of 24Recipient on the roster 24 of 24

Seats and moves 6 seats · 4 moves

Seats read off a team page, a governance page, an appointment notice or a filing that name this organization, and the dated moves into or out of it, from the observatory's own tables.

PersonSideSeat, as the record states itRoleDatesSourceId
Ajeya Cotraslice evaluatorBoard of Directors, Epoch AI (from its early-2025 spin-out)Board of Directors2025 – current, as readEpoch AI8f41b9b17e
Marius Hobbhahnslice evaluatorpart-time research fellow, Epoch, 2022 to 2023part-time research fellowformer2022 – 2023Hobbhahn CVebf0e84d58
Jaime Sevillaslice evaluatorCo-founder, CEO and board member, Epoch AI; Treasurer on its FY2024 990-EZCo-founder, CEO and board memberdates not statedEpoch AI; ProPublica; jaimesevilla.medcbd014ebf
Tamay Besirogluslice evaluatorCo-founder, Epoch AI; associate director to 2025, then research advisorCo-founderformer– 2025Epoch AI; Mechanize; his post35e2a3cf5e
Ege Erdilslice evaluatorResearcher, Epoch AI, left by Mar 2025Researcherformer– Mar 2025Epoch AI; Mechanize12a61628bc
Miles Brundageslice betweenStrategy advisor, Epoch AIStrategy advisordates not statedEpoch AI7284f229c3

Role read 6 of 6Organization read 6 of 6Organization profiled 6 of 6Dated 4 of 6

Moves

PersonFromToWhenSourceId
Ajeya CotrasliceCoefficient Giving , technical AI safety grantmaking; Redwood Research board (to Aug 2024)→METR technical staff; Epoch AI board2024 to 2026Coefficient Giving; METR; Epoch AI8a94650c89
Tamay BesiroglusliceEpoch AI (co-founder, associate director)→Mechanize (co-founder, Apr 2025); Google DeepMind (research scientist, by Sep 2026)2025 to 2026Epoch AI; Mechanize; his postd6d4d94fda
Ege ErdilsliceEpoch AI→Mechanize (co-founder)Apr 2025Mechanize8b99131748
Marius HobbhahnsliceEpoch (part-time research fellow)→Apollo Research (co-founder)2023Hobbhahn CVaf0df61432

From profiled 4 of 4To profiled 4 of 4Person profiled 4 of 4Dated 4 of 4

Filings