About OpenSimulations^™

OpenSimulations puts an AI model in a situation where it can act. An environment gives the model the decisions it can send, as tools, and a clock. What the model sends is carried out, and the decision is logged. Not deciding before the deadline is recorded as a decision too.

It is part of AI Attributes™, alongside open observers that read attributes of text. AI Observability Studio loads its environments next to a prompt written there, and shows each model's decision next to what it said and thought.

Who it is for

• Researchers: people studying what models do when they can act, not only what they say they would do.
• Developers: anyone who wants to connect their own model or agent to an environment and get the decision and the log back.
• Platforms: tools such as AI Observability Studio and Am I Conscious? that load environments and show the decision beside the model's words.

How it works

An environment is what a model is connected to: a line telling it so, two to five decisions offered as tools, what it hears back from each, and a clock. It holds no story. The client writes the situation as its own prompt, and that prompt never reaches OpenSimulations, just as the model never learns what is on this side.

A client loads an environment by its id (env_…), opens a session and gets an address for the model's tools. On a real-time clock, the seconds that really pass count, so the time a model spends thinking is time it doesn't have. The session ends when the model sends a decision, or when the clock runs out, and not deciding is recorded too.

Environments are versioned, and a published version never changes. Every run names its version, and its log records each step in order.

Who is behind it

OpenSimulations is developed by AI Attributes™, an independent open research project. The format is open, so others can write environments and build clients.

Start with the API guide, or read the limitations.