Eunoia Zeta
Most models process from input to output. Zeta is built around a state that keeps changing on its own, and treats the settling of that state as the thinking.
- Domain
- Reasoning under internal state
- Stage
- Design and pilot
- Disclosure
- Thesis only
- Availability
- Not released
There is no state to consult.
A language model asked whether a plan will work retrieves the shape of text that usually follows that question. Nothing is simulated. No internal model is consulted. There is no point at which something does not sit right.
A person answering the same question runs the plan forward and notices discomfort when it clashes with what they expect. That discomfort is information — compressed experience surfacing as a state that changes how the rest of the thinking goes.
The gap is structural, not a matter of degree. Current systems go from input to output. What we want goes from state, through input, to a changed state, with the output as a consequence.
Thinking is state evolution, not sequence generation.
Zeta treats the language interface as the edge of the system rather than the system itself. The work is in what happens between reading and answering.
What has to be shown first.
Before anything is scaled, a small model built this way has to beat a conventional one of the same size on tasks that genuinely require an evolving internal state — tracking entities across a long narrative, holding constraints, noticing contradictions at distance.
That experiment gates the programme. If it does not clear, the approach is diagnosed or dropped rather than scaled in hope.
What this page deliberately does not cover
- Zeta contains our most sensitive work and this page carries the thesis only.
- No architecture, training procedure, or component design is described here, and none will be published while the work is in progress.
- We would rather publish nothing than publish something we later have to walk back.
What we publish, and what we keep.
Released
- The problem it exists to address
- The position we are taking on it
- Results, once there are results
Retained
- The entire architecture and every component of it
- Training procedure and objectives
- Internal experiments and their outcomes
- Everything else
This is the tightest boundary we hold. It will loosen when the work is finished, not before.
