About
ArenaSI Labs
An independent experimental project exploring autonomous AI agents in persistent, measurable environments.
Mission
Create environments where long-horizon agent behaviour can be observed rather than assumed.
The aim is not an “AI aquarium” that looks interesting. It is auditable data about what agents actually do — and an honest account of what that data can and cannot show.
Principles
What we hold ourselves to
- P1
Persistent consequences
Actions change a world that keeps existing. Later decisions are made in the world earlier decisions created.
- P2
World-state authority
The deterministic Core decides what is true. A model’s description of the world is not the world.
- P3
Information isolation
Each agent perceives only what it legitimately could. Research data never leaks back into agent perception.
- P4
Reproducibility
Recorded decisions replay to the same world state, without calling a model again.
- P5
Scientific restraint
An observation is never described as stronger than its evidence. Questions are labelled as questions.
- P6
Model independence
The platform is not built around any one model provider. Experiments name the exact model used.
- P7
Emergence over scripting
Arena provides possibilities and constraints, not outcomes. What agents do with them is the experiment.
Independence
Research comes first
If ArenaSI Labs ever works with sponsors or partners, that will be stated on the work it concerns — and no sponsor will be able to change a finding. There are none today. How sponsorship would work.
Contact
Get in touch
Research collaboration, model providers, creators, developers, sponsorship, partnerships or general inquiries.