A non-profit AI Safety research organisation
Fundamental questions of safety and trust.
Understanding what safety claims establish.
Exploring how intelligent agents behave.
Making the evidence easier to see.
01 / Our mission
Evidence we can inspect.
Methods we can reproduce.
And room to be challenged.
We investigate AI safety from two directions: what safety claims actually establish, and how AI agents behave when they adapt and interact over time.
We build tools to trace claims to their assumptions and evidence, and controlled environments for testing explanations of agent behavior.
02 / Research programs
Different questions. A common commitment to traceable evidence and explicit limits.
What does a safety claim actually establish?
A searchable map connecting AI safety claims to their sources, assumptions, and evidence. Its current focus is debate and scalable oversight—helping researchers inspect a result, follow its dependencies, and identify what remains unresolved.
Explore the live Atlas
How does behavior change over time?
A research environment for studying agents that persist, interact, and adapt over time. Shared institutions, behavior records, and checkpoints help researchers investigate coordination, collusion, and responses to oversight.
Inside the program03 / Publications & outputs
Papers, preprints, technical reports, software, and datasets will appear here as they are released—with review status and limitations clearly stated.
No publications have been listed yet. Explore the questions guiding our first two programs below.
AI Safety Research Atlas / Research questions
AI Safety Research Atlas makes research easier to inspect: start with a claim, trace it to the original source, and see the assumptions and limitations that shape what it establishes.
The current pilot focuses on debate and scalable oversight, with linked papers, formal results, evidence grades, and open questions. AI-generated findings remain provisional until expert review.
A path through the Atlas
Multi-Agent Adaptive World / Research questions
How do agents coordinate across repeated interactions? When do shared institutions change incentives? How does behavior adapt to oversight?
Current results are bounded engineering demonstrations. Full scientific studies remain to be validated; a working environment alone does not establish an explanation of agent behavior.
04 / People & structure
Our team and advisors bring different perspectives to the questions at the heart of AI safety.
Our advisors bring external perspectives to our research.
How we organize the work
AI Safety Research Atlas and Multi-Agent Adaptive World are our first research programs. Further programs can grow from the questions our research reveals.
05 / Get involved
Review a claim. Test an assumption.
Help develop a reproducible study.
We welcome collaborators in formal methods, AI safety, multi-agent systems, experimental design, and research engineering.
Open roles will be posted here when available. We are currently shaping our team and research priorities.
No open roles listedCollaboration contact details will be added here soon.