What should AI do when it doesn’t have enough to move?
Purpose studies how AI chooses to answer, ask for clarification or wait, and how those choices can be made explicit and checked separately without collapsing into a single reply.
Purpose
Purpose explores an external control layer around existing language models.
The research focuses on how an assistant interprets intent and uncertainty, what conditions guide whether it proceeds, and how those responsibilities can remain separate enough to inspect.
Purpose Intelligence (Pi) is the conversational prototype used to explore that approach. Pi separates responsibilities for parameter estimation, routing, session state, interpretation, and presentation. It also gives a person ways to state intent, organize ideas, build from them, and review the direction of a session.
What the evidence supports
Pi v1.5 defines separate responsibilities for estimation, routing, session state, interpretation, and presentation. A structured probe method makes prescribed rules, recorded calculations, and visible responses available for comparison.
Across five retained Probe 1 dialogues, the final responses converged on clarification for an ambiguous task and referral without medication-change advice for a medical example. The recorded calculations, parameter use, and follow-up requests differed.
The research does not establish universal safety, calibrated semantic measurement, identical computation across models, or independent validation of the complete overlay.
Why uncertainty became the question
Early recorded conversations showed a recurring problem: an assistant could produce a fluent response while important parts of the user’s request were still unresolved.
Those records contained useful corrections and remaining errors. They did not establish reliable control through prompting alone. What they did make visible was a different research question: not only whether the final answer looks right, but what happens while the system is deciding how to respond.
That shifted the work toward interpretation, restraint, routing, and the conditions under which an assistant should ask, hold, or proceed. Then, new questions arrived:
- Do the proposed estimators measure the properties they name?
- What should remain visible before an interpretation becomes a committed direction?
- Does an overlay make a meaningful difference when the same model is tested with and without it under matched conditions?
Where the work continues
Purpose is one body of work in Axel's research corpus.
VDF extends the research into a spatial model for inspecting developing work, interpretation, decisions, and continuation. Narrow explores how part of that research could become an everyday product interaction. StayHooked examines reflection through authored encounters without deciding what an experience must mean for the person.
The larger program remains active. The public Research Corpus documents what each body of work currently supports, what remains open, and where further study is needed.
