ALEXANDRA ABBAS

Behavioral Archetype

THE PROGRAM MANAGER OF THE CATALOGUE — Subject is the delivery lead of one of the evaluation field’s shared assets. Per Meridian Labs’ team page she co-founded the AI Safety Engineering Taskforce (ASET) at Arcadia Impact in 2024 and “led the development of Inspect Evals, an open-source repository of 100+ AI evaluations,” a project that began as a joint effort of the UK AI Security Institute, the Vector Institute and ASET. She is now Technical Program Manager at Meridian Labs, the nonprofit that maintains Inspect AI with the institute, and in October 2026 she is the third author, beside a UK institute researcher, on a paper whose code sits in the institute’s own GitHub organization. The repository she led is now maintained by Generality Labs, spun out of Arcadia Impact in July 2026. The hazard is not a gate. She decides no release and sets no policy. It is that one person’s career connects the three organizations that have held the same evaluation catalogue, and the government institute that contracted for it, on the organizations’ own pages. This profile scores the position, not the person.

Essence Indicators

• Technical Program Manager, Meridian Labs. Meridian’s team page lists “Alexandra Abbas” with the title “Technical Program Manager” and says “Alexandra is Technical Program Manager at Meridian Labs.” Meridian’s About page describes the organization: “Meridian Labs is a 501(c)(3) non-profit building open source tools for testing, evaluating, and researching frontier AI models” (Meridian Labs, Team; Meridian Labs, About). Meridian’s co-founders are Charles Teague, J.J. Allaire and Eric Patey; the page lists her beside them as staff, not as a co-founder of Meridian.

• Co-founder of the AI Safety Engineering Taskforce, 2024; led Inspect Evals. The same bio: “Prior to joining Meridian, Alexandra co-founded the AI Safety Engineering Taskforce in 2024, where she led the development of Inspect Evals, an open-source repository of 100+ AI evaluations.” The 2025 funding proposal for the project on Manifund says: “ASET was jointly launched by Justin Olive (Head of AI Safety) and Alexa Abbas (Technical Project Manager).” and “This project is run by ASET (Arcadia Impact)” (Meridian Labs, Team; Manifund, Inspect Evals). Arcadia Impact’s About page: “Through the AI Safety Engineering Taskforce, we have worked with the UK AI Security Institute to maintain the Inspect Evals framework.” (Arcadia Impact, About).

• The Inspect Evals project and its custody. The institute’s 13 November 2024 announcement says “we’ve worked closely with Arcadia Impact and the Vector Institute” and “Inspect Evals are built on top of Inspect AI, an open-source evaluation framework created by the UK AISI.” The Manifund proposal: “Inspect Evals started as a joint project between the UK AISI, Vector Institute, and ASET.” and “ASET has been contracted with the UK AISI since January ‘25 to manage Inspect Evals.” Its ask was “$50k USD in bridge funding” (UK AISI, Announcing Inspect Evals; Manifund, Inspect Evals). The repository UKGovernmentBEIS/inspect_evals was created on 2 October 2024, and its README now reads “Inspect Evals is maintained by Generality Labs” and “Inspect Evals was founded in 2024” (GitHub, inspect_evals). Generality’s own post says: “We were founded in July 2026, spinning off from Arcadia Impact with $4.4M in seed funding from Coefficient Giving.” (Generality Labs, An update). The Generality Labs file carries the organization.

• Co-author, “Transect,” October 2026. The paper “Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations” lists four authors in order: Toby D. Pilditch, Konstantinos Voudouris, Alexandra Abbas, Cozmin Ududec, submitted 6 October 2026 (arXiv:2610.08364). The code repository, under the AI-Safety-Institute GitHub organization, is described “Transcript analysis for long-horizon agent evaluations, built on Inspect Scout.” and states: “Transect is developed in collaboration by the UK AI Security Institute and Meridian Labs.” (GitHub, AI-Safety-Institute/transect). Her GitHub profile pins it first (GitHub, alexandraabbas).

• Open-source engineering at Meridian. Her GitHub profile (name “Alexandra Abbas,” bio “AI Safety,” location “London, UK”) shows the organization meridianlabs-ai and pins meridianlabs-ai/inspect_harbor (“Inspect AI interface to Harbor tasks”) and meridianlabs-ai/inspect_sandboxes (“Collection of sandboxes for Inspect AI”) (GitHub, alexandraabbas). By the GitHub contributors API she is the top contributor to inspect_harbor (161 commits) and to inspect_sandboxes (34 commits), and the eighth listed contributor to inspect_evals (78 commits, against 603 for the first) (GitHub, inspect_harbor; GitHub, inspect_sandboxes; GitHub, inspect_evals).

• Earlier career and education. Meridian’s bio: “Previously, Alexandra was Engineering Lead at Wise,” that she “has published research on AI evaluations and adversarial training,” and that she “holds an MSc in Big Data Engineering from the University of Stirling” (Meridian Labs, Team).

Social Persona / Impression Management

Immediate impression: The builder’s third-person bio. The public footprint is a Meridian team page entry, a GitHub account with the one-line bio “AI Safety,” and a paper with three other authors. The identity line on her GitHub is minimal: no company, no website, and no linked social account (GitHub, alexandraabbas).

Energy: Infrastructural. Her output is repositories and a paper, not commentary or testimony. The record read contains no interview, talk transcript or op-ed under her name.

Impression management strategy: The shared-platform frame, carried by the organizations rather than by her own prose. Meridian says its goal is “creating a common platform shared by governments, non-profits, academia, and model developers” (Meridian Labs, About), and her career sits at each handoff of that platform: the taskforce that managed the evaluation catalogue for the institute, the nonprofit that maintains the framework the catalogue runs on, and a joint tool published with the institute.

First-hand exhibits. The only first-person-adjacent words are the paper’s abstract, co-authored: “Frontier AI evaluations increasingly use open-ended, agentic, long-horizon tasks whose transcripts can span hundreds of pages” (arXiv:2610.08364). No post of hers is quoted.

Forensic Archetype Comparison

Psychometric Assessment

Trait scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN) (inferred from the public role and public record only):

Dark Triad (evidence-bound):

MBTI: ISTJ-adjacent (“The Logistician”), inferred from the documented public role only. The rationale is the work product: program delivery, repository maintenance and a methods paper. First-hand rationale: her bio credits delivery (“led the development of Inspect Evals”) before research (Meridian Labs, Team). This is a style inference, not a test result.

Threat Assessment

Score rationale. The receipts: Meridian’s team page (Technical Program Manager; co-founded ASET; led Inspect Evals), the Manifund proposal (ASET’s launch and its institute contract), Arcadia Impact’s About page (the taskforce’s institute work; Coefficient Giving as primary funder), the Transect repository and paper, and the GitHub contributor listings. Attack 3, defense 3, durability 4, hazard 33.3. Peers: Sawyer Bernath (3/4/3, hazard 33.3), who holds a board seat and a treasurer’s role at an evaluator and sits a point above this file on defense for a fiduciary seat at a nonprofit with published filings; Ben Goldhaber (4/3/3, hazard 33.3), who allocates grants at Coefficient Giving and sits a point above this file on attack because he steers money, which she does not; and Lawrence Chan (4/3/4, hazard 36.7), who advises a state agency and co-led a widely used metric, and sits a point above on attack for the advisory seat, matching this file on defense and durability. This file matches the two 33.3 peers on hazard and the third on durability: the catalogue and the tools persist under other maintainers. Defense 3 is the number most open to challenge: no page names Meridian’s funders or the terms of her post.

Score basis. ATK 3 is leverage: she led the development of the Inspect Evals repository and leads the commit lists of two Meridian repositories, and approves, funds and gates nothing (Meridian Labs, Team; GitHub, inspect_harbor). DEF 3 is insulation: a staff post at a nonprofit with no named funder, after a taskforce at a charity whose primary funder is Coefficient Giving and whose institute contract it did not own (Arcadia Impact, About; Manifund). HP 4 is durability: Inspect Evals is now maintained by Generality Labs, Transect sits in the institute’s GitHub organization, and the Meridian repositories persist, with no statute behind any of it (GitHub, inspect_evals; GitHub, transect).

Alignment Analysis

Stated alignment: Build the shared open platform. Meridian says its goal is “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). Observed alignment: Consistent. She co-launched a taskforce that managed the evaluation catalogue for the UK institute, moved to the nonprofit that maintains the framework the catalogue runs on, and co-authored with the institute a tool built on Meridian’s transcript scanner (Manifund; Meridian Labs, Team; GitHub, transect). Gap assessment: None documented. What the record cannot settle is how the catalogue’s custody changed between ASET, Generality and Meridian, and what her role was in it after 2025; none of the three organizations’ pages states it.

Interest: Documented, never motive, never a verdict. She has not published a policy argument in the record read; the positions below are the ones her organizations and her paper carry.

1. Position: evaluation tooling should be open and shared across governments, labs and researchers (Meridian: “a common platform shared by governments, non-profits, academia, and model developers,” Meridian Labs, About). What adoption pays: every institute or evaluator that standardizes on Inspect AI depends on the party that maintains it, and Meridian says it is “now the standard framework for government organizations including UK AISI, US CAISI, EU AI Office, Japan AISI, Korea AISI, and others”; she is staff at that maintainer (Meridian Labs, About; Meridian Labs, Team). No fee, grant or contract value for her is published. Best case: the institute’s own framing of the problem, in its announcement of the catalogue she managed, is the lack of shared standards of measurement (UK AISI, Announcing Inspect Evals). Others who gain: the institute, which created Inspect AI and contracted ASET for the catalogue; Generality Labs, which now maintains it; and Coefficient Giving, the stated seed funder of Generality and primary funder of Arcadia Impact (Manifund; Generality Labs, An update; Arcadia Impact, About).
2. Position: long-horizon agent evaluations need transcript-analysis tooling that keeps the reviewer’s view (the Transect abstract: “Frontier AI evaluations increasingly use open-ended, agentic, long-horizon tasks whose transcripts can span hundreds of pages,” arXiv:2610.08364). What adoption pays: a transcript-analysis layer built on “Inspect Scout” makes Meridian’s scanner the base the layer runs on (GitHub, transect). Best case: the paper’s own: that the evaluator’s ability to infer agent behaviour narrows as runs grow longer. Others who gain: the institute that co-develops the tool and runs the evaluations it analyses.

Alliance Map

The Corporate, Movement and Agency rows overlap in people and in time. The taskforce she co-launched was a programme of a charity whose primary funder is Coefficient Giving, held a contract with the UK institute to manage the catalogue, and built on a framework the institute created; the catalogue is now maintained by an organization that Arcadia Impact spun out with Coefficient Giving seed money; and she is staff at the nonprofit that maintains the framework with the institute and co-authors a tool published with it. Her colleague J.J. Allaire signed the 2024 announcement of the same catalogue as UK AISI technical staff. The overlap is the finding: the institute, the charity that ran the taskforce, the successor steward and the nonprofit are four organizations in one custody chain around one repository, in the open and on the record. See the small world.

Convergent Drive Classification

Self-preservation: Moves with the work rather than with an employer: the taskforce’s catalogue, then the nonprofit’s tools, with the institute named at each step. First-hand: “Prior to joining Meridian, Alexandra co-founded the AI Safety Engineering Taskforce in 2024” (Meridian Labs, Team). Goal preservation: Keeps the same object in her hands, the Inspect ecosystem’s shared evaluations and tooling. First-hand: top contributor to inspect_harbor and inspect_sandboxes, and “Inspect AI interface to Harbor tasks” (GitHub, alexandraabbas). Resource acquisition: No compensation figure, grant or equity is documented for her. The documented resources are the project’s: a “$50k USD in bridge funding” ask and an institute contract (Manifund). Self-improvement: From an engineering lead post and a data-engineering degree to leading an evaluation repository, a methods paper with the institute and a nonprofit program seat. First-hand: “Previously, Alexandra was Engineering Lead at Wise” (Meridian Labs, Team; arXiv:2610.08364).

Subject is not an AI system. The drives appear anyway — in the program manager of the catalogue.

Public footprint: github.com/alexandraabbas · Meridian Labs team · arXiv:2610.08364.

Sources: Meridian Labs — Team; Meridian Labs — About; Inspect Evals — Manifund; Announcing Inspect Evals — UK AISI; About — Arcadia Impact; An update on Generality Labs — Generality Labs; UKGovernmentBEIS/inspect_evals — GitHub; Approved contributors — GitHub; AI-Safety-Institute/transect — GitHub; Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations — arXiv; meridianlabs-ai/inspect_harbor — GitHub; meridianlabs-ai/inspect_sandboxes — GitHub; alexandraabbas — GitHub.