Control while work is moving.
Evaluate and constrain actions at the point where an agent wants to act.
Research / Runtime assurance
UBIKITOS studies what it takes for autonomous systems to act usefully inside explicit boundaries: runtime assurance, supervisory control, provenance, observability, uncertainty and epistemic limits.
A new generation
Live: R3, including R3.9. The WorkMachine's third-generation milestone is here.
In development: R4.17. The Untrusted Output Boundary anchors ongoing native integration: signed authorization, guarded results and recovery. Independent review and formal acceptance remain open.
Explore the release journey ↗01 / The question
The central research question is simple to state and difficult to operationalise: when does the available observation regime fail to justify consequential action—and what should happen then?
UBIKITOS treats a confident answer as a proposal, not a permission slip. The research boundary asks what the system can do, what supports the action, who has authorised it and what changes if the decision is wrong.
02 / Research themes
Evaluate and constrain actions at the point where an agent wants to act.
Define when a human should approve, review, correct or stop the workflow.
Preserve where evidence came from, when it was observed and what it supports.
Distinguish missing evidence and unknown conditions from low model confidence.
Do not let a fluent proposal silently outrun the observation regime around it.
03 / Experimental systems
UBIKITOS develops small, inspectable experiments rather than treating autonomy as a single benchmark. Each experiment makes the boundary explicit, records what happened and asks whether the control was useful enough to keep.
Research standard: a claim should be paired with a trace, a failure mode or a measured before-and-after—not only a more persuasive interface.
04 / Continue