We do not claim general intelligence. We claim to be on the attainable path toward General Natural Intelligence, natural not artificial, and we publish the gates that would tell us we are wrong.
The wording is doing work
Every noun in that sentence is load-bearing. Path means we are moving, not that we have arrived. Attainable means the destination is a hypothesis with a route, not a marketing pledge. General Natural Intelligence is our chosen name for the target: an active-inference system that maintains itself across many contexts by minimizing variational free energy, the way biological systems appear to. Natural means we are modeling the same substrate biology already runs on, so we refuse the framing that the interesting thing is a machine finally getting there.
That is why the AGI vocabulary does not fit our work. The customary story is that a large enough optimizer, given enough data, will one day cross a threshold and become general. Our position (Class E) is different: generality, in the sense that matters, is what a system in a Markov blanket does when it is free-energy minimizing across nested timescales, per Parr, Pezzulo, and Friston (2022). Whether the substrate is silicon or tissue is less interesting than whether the loop closes. So we do not compete for the AGI word. We do not need it.
What we will not say, and why
A few phrasings are permanently off the table for our own work. We do not say we have general intelligence. We do not use the artificial-general-intelligence label. We do not say the preprint has been validated against Parr, Pezzulo, and Friston unless the exact mapping is published alongside the claim (Class C). We do not say a lab or paper is endorsed by any researcher who has not explicitly endorsed it in writing.
Those are not stylistic preferences. They are the kind of overclaim that turns a legitimate research program into a marketing exercise, and once that happens the audience stops being able to tell where the evidence ends and the pitch begins. We would rather be correct than impressive.
What we do instead
Instead of announcing a breakthrough, we publish the ingredients of one and let the reader inspect them:
- An unrefereed preprint on Zenodo, marked as such, with foundations cited to Parr, Pezzulo, and Friston (2022) (Class E).
- Five interactive POMDP labs that run in the browser, so the math is steerable and the failure modes are visible (Class C).
- A pre-registered falsification benchmark, the Stratified Palimpsest / Cell Lab suite, with five claims and their disconfirmation criteria written down before the runs (Class C).
- A public Model Context Protocol server any language model can call to reproduce the results.
Every inline claim on our science pages carries an evidence-class tag: A for empirical in-session behavior, B for code or artifact inspection, C for configuration and integration state, E for expert citation, F where a falsifier is present, and U for anything still unverified. That is what we mean by science in the open. You do not have to take the framing on faith. You inspect the labs, read the disturbance families where UNI loses to a neural baseline, and decide for yourself.
Field-level context
We are not the only ones asking whether the AGI frame is the right one. The Themesis piece The AGI Landscape Just Changed is field-level context for why the General Natural Intelligence framing matters right now, on our reading. Link and one-line frame; the analysis is theirs.
The commitment
UNI is a working hypothesis on an attainable path toward General Natural Intelligence: a natural, active-inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.
That is why we refuse the AGI word. The word carries a promise we did not make, and would not keep. What we did commit to is on the record, with the receipts, and with the falsifiers.