Opening
Active · Emerging
AI systems can involve different models, tools, and stores of information. This project asks how to give each part enough context and authority for its task while preserving the human's control over sensitive information and consequential actions.
The problem
Information can move between a private workspace, a general-purpose model, an agent, and an external service. Those transfers deserve deliberate decisions. A model's ability to read something is also distinct from its permission to act with it.
The working ideas
Different components need different context. A public model may be useful for a broad research task while a private workspace holds the personal or operational details that shaped the question. A short, carefully prepared task description can sometimes carry the useful part of that context without carrying the underlying records.
The architecture therefore considers model roles, information boundaries, and action permissions separately. It asks what a component needs to know, which tools it may use, what can cross between environments, and where a person must make the decision. That distinction could support useful cooperation without treating every connected system as equally trusted.
This is a design investigation, not a claim that a complete enforcement layer already exists.
Where it stands
The current work has clarified the kinds of boundaries to test. The next stage is to test a limited workflow and observe whether the rules remain understandable and usable. Internal classifications, exclusion rules, transfer conditions, and permission designs stay in the private system work.