Research that has to operate.
We study artificial intelligence for long-horizon work: systems that can be inspected, that continue under constraint, and that leave judgment with people.
Systems for work
that takes days.
Long-horizon work. Adaptive systems. Judgment that stays with people.
How we build
A method for complex-task AI. Not a model card, and not a reading list.
The task is the unit
We do not begin with a model and search for a use. We begin with a task that must exist, then build the intelligence it requires.
Make the system inspectable
If we cannot see the state, the action, and the rationale, we cannot hold the work. Architecture over fashion.
Prefer reliable machinery
World, plan, exception, handoff. The unglamorous loop is the product. Spectacle is a cost.
Reserve judgment for people
When the system meets a decision that is judgment rather than computation, it stops and asks. That is the design, not a failure.
Inquire about the work.
Partners and candidates. A brief note is sufficient.