MERIDIAN LABS

Institutional Archetype

THE MAINTAINER OF THE MEASURING STICK — Meridian Labs does not publish a ranking, a benchmark score or a finding about a company. It maintains the software that other people use to produce them. Its flagship, Inspect AI, is “an evaluation framework built by Meridian’s founding team in collaboration with the UK AISI” (Meridian Labs, About), and is the harness the UK AI Security Institute uses; the US Center for AI Standards and Innovation ran a benchmark in it; METR is moving its tooling onto it. Since 7 May 2026 it also maintains Petri, the alignment auditor Anthropic built and gave away, and Anthropic’s own system card now reports Petri runs “conducted by Meridian and the UK AI Security Institute.” The tool that grades the models is maintained by a nonprofit founded in 2025 whose own pages name nobody who pays for it.

Mandate & Origin

“Meridian Labs is a 501(c)(3) non-profit building an open source platform for understanding, evaluating, and testing AI models and agents” (Meridian Labs). Its About page gives the goal: “to advance AI safety, alignment, and security work by creating a common platform shared by governments, non-profits, academia, and model developers” (Meridian Labs, About).

Its team page lists Charles Teague (CEO, co-founder), J.J. Allaire (Principal Engineer, co-founder) and Eric Patey (Principal Engineer, co-founder), who “co-founded Meridian Labs in 2025,” with Alexandra Abbas (Technical Program Manager), Ransom Richardson, Matthew Brandly and Kai Fronsdal (Meridian Labs, Team). Teague “was a Technology and Security Policy Fellow at the RAND Corporation, where he helped develop Inspect AI”; Allaire “is also a visiting researcher at the US CAISI, focusing on platform technology, cybersecurity, and AI measurement science”; Fronsdal “was an Anthropic Fellow and MATS AI Safety Researcher, where he developed Petri” (Meridian Labs, Team). Allaire’s name appears earlier on the UK institute’s own site: the 13 November 2024 post announcing Inspect Evals is signed “JJ Allaire, Technical Staff” (UK AISI, Announcing Inspect Evals).

The legal form is on the IRS record, not on Meridian’s pages: Meridian Research Labs Inc, EIN 33-3145457, in care of Charles Teague at 24 School Street, Boston, a 501(c)(3) with the exemption issued in March 2025 (ProPublica Nonprofit Explorer). Its pages give the street address as 501 Boylston St, 10th Floor, Boston (Meridian Labs). Its GitHub organization, meridianlabs-ai, was created on 25 January 2025 and holds 27 public repositories (GitHub).

Funding & Backers

No funder is named. The home, About, Team, Projects, Careers and Blog pages state none; Anthropic’s donation post gives no financial term; the Petri 3.0 acknowledgements thank Anthropic “for entrusting its development to us” and the UK institute “for their continued collaboration on this work,” and name no grant (Meridian Labs, Introducing Petri 3.0; Anthropic, donation post).

• The filing. ProPublica lists the organization with “No Financial Data Available”: the exemption dates from March 2025 and no Form 990 has been digitized (ProPublica Nonprofit Explorer).
• The payroll it advertises. Open roles are Member of Technical Staff, Engineering and Research (“Senior and principal levels · $175,000 – $325,000”) and a Technical Operations Manager ("$110,000 – $160,000"), with health, dental and vision cover (Meridian Labs, Careers). A seven-person staff at those bands implies a budget in seven figures; no page states one.
• Grant funders. Neither the Survival and Flourishing Fund’s 2024, 2025 and 2026 recommendation tables nor Coefficient Giving’s grant search lists Meridian Labs or Meridian Research Labs (SFF 2024; SFF 2025; SFF 2026; Coefficient Giving, grants search). Generality Labs, which maintains Inspect Evals and shares Meridian’s lineage through the AI Safety Engineering Taskforce (an Arcadia Impact programme that Abbas co-founded), states “$4.4M of seed funding from Coefficient Giving” (Generality Labs); Meridian states nothing comparable. See The EA to AI-Safety Funding Pipeline.

Institutional Voice & Intent

Meridian’s account of itself: “Inspect AI is now the standard framework for government organizations including UK AISI, US CAISI, EU AI Office, Japan AISI, Korea AISI, and others. It is also the standard framework for research organizations including METR, Apollo, Epoch, SecureBio, Redwood, and RAND” (Meridian Labs, About). Its Careers page: “Inspect is in use at AI Safety Institutes in the UK, US, EU, Japan, and Korea, and at research organizations including METR, Apollo, Epoch, SecureBio, and RAND” (Meridian Labs, Careers). Its closing line: “As AI capabilities advance, the evaluation and research infrastructure surrounding them must keep pace. Meridian exists to make sure it does” (Meridian Labs, About).

Stated intent: Provide a common open-source platform for governments, nonprofits, academia and developers to test frontier models, and “providing the infrastructure for AI Safety Institutes, research organizations, and lab safety teams to transfer their work into broadly-available open source projects” (Meridian Labs, About).

Observed intent: Take custody of tools that institutes and labs would rather not own. The UK institute created Inspect and Meridian’s founders built it; Anthropic’s Petri and Bloom moved to Meridian; Meridian then runs evaluations itself with those tools, which its About page lists as “Conducting pre-deployment testing using our tools,” and its Careers page confirms: “Petri alignment audits by our team appear in recent Anthropic model cards.”

Gap: The structure is stated, not hidden. Inspect’s repository still sits under the UK government’s GitHub organization with Meridian listed beside the institute in the docs; the funder of the steward is the unstated part. What the record cannot settle is whether the nonprofit’s priorities and its users’ would diverge, because no page says who sets Meridian’s budget. For a maintainer of public tools, a published funder list would settle it; the pages do not give one.

Interest: Meridian argues for a shared open-source evaluation stack maintained by an independent steward. What adoption would pay it is set out from its own pages; no fee, contract or grant value is published.

• One common platform, maintained by Meridian. Position: “a common platform shared by governments, non-profits, academia, and model developers” (Meridian Labs, About). What it pays: every institute or lab that standardizes on Inspect, Scout, Flow or Petri depends on the party that maintains them, and Meridian’s seven-person payroll and open roles are its only disclosed cost (Meridian Labs, Careers). Best case, its own: “As AI capabilities advance, the evaluation and research infrastructure surrounding them must keep pace.” Convergence: the UK institute offloads the upkeep of a tool it created while its researchers remain Meridian co-authors; METR and Apollo replace in-house frameworks (UK AISI, Engineering Playbook).
• Independence of the auditor from the lab that built it. Position, Anthropic’s: handing Petri over “will help ensure that Petri remains independent of any AI lab, so that its results will be seen as neutral and credible by those across the industry and beyond” (Anthropic, donation post). What it pays: Meridian gets the tool and its users; Anthropic gets a neutral-credible runner for the cross-developer comparison in its own system card, and “Anthropic will continue to support Petri and use it in its own alignment assessments” (Meridian Labs, Introducing Petri 3.0). Best case, Anthropic’s: the MCP donation to the Linux Foundation is the precedent it cites. Convergence: the lab, the UK institute and the nonprofit each gain from the same arrangement, and the buyer of the neutrality is the reader of the card.

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

Position in the Apparatus

• Inspect, the institutes’ harness. The UK institute: “In May 2024, the AI Security Institute open-sourced” Inspect, which “powers many of AISI’s frontier evaluations”; “METR migrated from their in-house Vivaria platform to Inspect,” “Apollo Research deprecated their internal evaluation framework after a three-month trial of Inspect,” and “Other government institutions have adopted Inspect and the sandboxing toolkit for their own evaluation work”; Inspect has “240+ contributors to inspect_ai alone” (UK AISI, Releasing AISI’s Engineering Playbook, 18 June 2026). The docs call Inspect “a framework for frontier AI evaluations developed by the UK AI Security Institute and Meridian Labs” (Inspect). The repository, MIT licensed, is UKGovernmentBEIS/inspect_ai, with the account jjallaire first on its contributor list at 2,850 contributions (GitHub).
• The US institute and METR. CAISI’s evaluation of DeepSeek models: “including modifying the benchmark to run in the Inspect evaluation framework,” with a footnote to inspect.aisi.org.uk (NIST CAISI, Evaluation of DeepSeek AI Models). METR’s Vivaria repository: “METR is transitioning its internal tooling from Vivaria to Inspect” (GitHub, METR/vivaria). Meridian lists Scout “in use at UK AISI, US CAISI, METR, and Apollo” and Flow “in use at UK AISI and US CAISI” (Meridian Labs, About). See SecureBio, whose evaluation code runs on Inspect, and Apollo Research.
• Petri changes hands (7 May 2026). Anthropic: “We have handed over its development to Meridian Labs, an AI evaluation nonprofit”; “Petri has been part of our alignment assessment for every Claude model since Claude Sonnet 4.5”; the UK institute “made it a major part of how they evaluate models for their propensity to sabotage AI research” (Anthropic, Donating our open-source alignment tool). Meridian’s Petri 3.0 post, co-written by Meridian’s Fronsdal, Allaire and Teague with three UK institute researchers and a Cambridge Boston Alignment Initiative author, says the UK institute “built its alignment evaluation pipeline on Petri” and “used a prototype of 3.0 in its pre-deployment evaluations of Claude Mythos and Opus 4.7,” and that Bloom “has moved to Meridian alongside Petri” (Meridian Labs, Introducing Petri 3.0). The Projects page lists Petri, Petri Bloom, Petri Dish, Inspect Flow, Inspect SWE, Inspect Harbor, Inspect Viz, Inspect Sandboxes and Inspect VSCode (Meridian Labs, Projects).
• In the lab’s system card (9 June 2026). Anthropic’s Claude Fable 5 and Claude Mythos 5 system card: “we have also released the open-source package Petri, now managed and maintained by the independent nonprofit Meridian Labs,” and “For comparison with models by other developers, we report results from a set of Petri 3.0 evaluations conducted by Meridian and the UK AI Security Institute (UK AISI).” The run “used Claude Opus 4.7 and GPT-5.5 as investigator models and Opus 4.7, GPT-5.5, and Gemini 3.1 Pro as judge models,” and the card concludes that Mythos 5 “appears to be tied with other recent Claude models as the best-aligned publicly accessible model on nearly all these metrics” and “better than competing models” (Anthropic, Claude Fable 5 and Mythos 5 system card). The cross-developer comparison in a lab’s release document was run by the nonprofit that lab donated the tool to, and Anthropic’s own Opus 4.7 is among its investigators and judges.

Actions & Leadership Choices

CONDUCT: WATCH — INFRASTRUCTURE WITH NO NAMED FUNDER. No wrongdoing is alleged. Meridian maintains public, MIT-licensed tools and its pages say what each is for and who uses it. The structural point is the one every harness shares: an evaluation is only as neutral as the code and the seat that runs it, and for Inspect and Petri that seat is a seven-person nonprofit that states its mission and its users and not its money. The same institutes that grade labs depend on its tools, and the lab that gave it Petri reports, in its own card, Petri runs conducted by the nonprofit.

• 14 November 2023 to May 2024. The Inspect repository is created and the UK institute open-sources it (GitHub; UK AISI, Engineering Playbook).
• 13 November 2024. The UK institute announces Inspect Evals in a post signed by Allaire as Technical Staff (UK AISI).
• 25 January 2025. The meridianlabs-ai GitHub organization is created (GitHub).
• March 2025. The IRS exemption for Meridian Research Labs Inc is issued (ProPublica Nonprofit Explorer).
• 7 September 2025. Inspect Viz is introduced; Inspect Scout (25 February 2026), Harbor Tasks (12 February 2026), Inspect Flow (4 March 2026), Long-Horizon Agents (26 April 2026) and Cloud Sandboxes (30 April 2026) follow (Meridian Labs, Blog).
• September 2025. The US institute’s DeepSeek evaluation runs a benchmark in Inspect (NIST CAISI).
• 7 May 2026. Anthropic hands Petri to Meridian; Petri 3.0 and the move of Bloom are announced the same day (Anthropic; Meridian Labs).
• 9 June 2026. Anthropic’s Fable 5 and Mythos 5 system card reports Petri 3.0 evaluations conducted by Meridian and the UK institute (Anthropic).
• 18 June 2026. The UK institute’s Engineering Playbook names METR, Apollo and “other government institutions” as Inspect users (UK AISI).

Alliance Map

Where ties on different axes run at the same time, the overlap is the finding. The lab that built Petri gave it to the nonprofit; the nonprofit’s founders built Inspect with, and in two cases worked for, the UK institute that now runs it; the US institute runs it too and hosts a Meridian co-founder as a visiting researcher; and the lab’s own system card reports the cross-developer comparison run by the nonprofit and the UK institute together. The documents show disclosed collaboration, not coordination; the one thing they do not show is who pays Meridian.

Reach Assessment

Institutional: The default evaluation harness at the UK institute and, by its own statement and the US institute’s DeepSeek report, at the US institute; the maintainer of a lab’s alignment auditor. Memetic: Low; Meridian publishes no scores of its own, and its name reaches readers through a lab’s system card, in one section. Civilizational: If evaluation is the layer that tells governments and labs what a model does, the people who maintain the harness sit upstream of every result run in it. Nothing in the record says Meridian has used that position for anything but upkeep, and the code is open and MIT licensed, so a user can read what a score computes. The gap is institutional: the steward is three co-founders and a small staff, its funding is unstated, and the lab that donated the auditor reports runs of it in its own release document. A published funder list and an audited budget would close most of that distance.

Score basis. Atk 8 is reach: Meridian’s About page states that Inspect is the standard framework at UK AISI, US CAISI, the EU AI Office, Japan AISI and Korea AISI and at METR, Apollo, Epoch, SecureBio, Redwood and RAND; CAISI’s DeepSeek evaluation ran a benchmark in Inspect; METR’s repository says it is moving to Inspect; and Anthropic’s Fable 5 and Mythos 5 card reports Petri 3.0 evaluations “conducted by Meridian and the UK AI Security Institute” (Meridian Labs, About; NIST CAISI; GitHub, METR/vivaria; Anthropic, system card). The points under 9 are that no statute or procurement standard names it and that the lab card cites it for one evaluation among many. Peers on atk 8: Apollo Research and Gray Swan; the UK AI Security Institute, METR and CAISI hold 9. Def 5 is funding insulation: no page names a funder, a budget or a donor, and the IRS record shows a March 2025 exemption and no Form 990 data (ProPublica Nonprofit Explorer). Nothing is disclosed to weigh, in either direction: the number reflects an unknown, not evidence of dependence. Peers on def 5: Andon Labs; Gray Swan, FAR.AI and Palisade Research hold 6 on a published cap, split or philanthropic base. Hp 7 is entrenchment: Inspect’s repository sits under a UK government GitHub organization, the UK institute’s playbook counts 240+ contributors and names METR and Apollo as users, and Meridian now maintains Petri, Bloom, Scout and Flow (UK AISI, Engineering Playbook; GitHub; Meridian Labs, Projects). The points off are that no statute installs it, that Meridian was founded in 2025, that it has filed no Form 990 and that no multi-year contract is disclosed. Peers on hp 7: Epoch AI and Gray Swan; FAR.AI, Palisade Research and Transluce hold 6.

Sources: Meridian Labs — home; About — Meridian Labs; Team — Meridian Labs; Projects — Meridian Labs; Careers — Meridian Labs; Meridian Labs Blog; Introducing Petri 3.0 — Meridian Labs; Donating our open-source alignment tool — Anthropic; System Card: Claude Fable 5 and Claude Mythos 5 — Anthropic; Inspect — UK AI Security Institute and Meridian Labs; UKGovernmentBEIS/inspect_ai — GitHub; meridianlabs-ai — GitHub; Announcing Inspect Evals — UK AISI; Releasing AISI’s Engineering Playbook — UK AISI; Evaluation of DeepSeek AI Models — NIST CAISI; Vivaria — METR, GitHub; Generality Labs; ProPublica Nonprofit Explorer — Meridian Research Labs Inc; SFF 2024; SFF 2025; SFF 2026; Coefficient Giving — grants search.