Every public sentence about UNI is meant to be inspected. This guide gives you the parts to look for and the shape they take, so you can decide for yourself what to believe.
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.
What a UNI claim actually contains
A well-formed UNI claim carries four pieces. If any of them is missing, that is a signal to slow down before you accept it.
- The proposition. One sentence stating what is asserted. Not a mood, not a slogan, an actual statement that could be right or wrong.
- The evidence class. A short tag such as Class A or Class C that tells you what kind of thing is backing the claim. Empirical observation, code inspection, integration state, expert citation, or explicit unverified.
- The falsifier. The condition that would show the claim is wrong. If a claim has no falsifier, it is a mood, not a claim.
- The ledger entry. A pointer to the append-only record where the observation was written down, dated, and left where it can be read again.
The evidence classes, briefly
These tags do a lot of work. They are the honest label on the tin.
- Class A empirical, observed in a live session. The dial moved, the output changed, and someone watched it happen.
- Class B code or configuration inspection. The behavior is what the source says it is, not what the running system was seen doing.
- Class C integration or configuration state. Two components are wired the way the claim says they are wired.
- Class E expert citation. Someone qualified said this in a citable place. Naming a source is not the same as running the experiment.
- Class F falsifier present. The claim ships with the condition that would refute it.
- Class U unverified. Stated plainly, not smuggled in.
Two examples make the difference concrete. "The Precision Lab's sensory precision dial produces the behavioral regimes described in Parr, Pezzulo and Friston (2022)" is Class E: it rests on a citation. "The Cell Lab's active-inference controller lost to the neural baseline on memory_leak and cpu_noisy_neighbor" is Class A: it is a recorded outcome from a pre-registered run. Both are useful. They are not the same thing.
How to read a claim in five steps
- Find the proposition and read it slowly. If it is fuzzy, ask what would count as it being true.
- Find the evidence-class tag. If none is present, treat the sentence as Class U until you find the receipt.
- Find the falsifier. If the claim cannot fail, it is not carrying information.
- Follow the pointer to the ledger entry. Check the date, the source, and whether the entry has been superseded.
- Ask what the claim does not yet prove. A passing test is not a working feature. A cited paper is not a run.
What a UNI claim will never do
The public copy is fenced. UNI does not claim general intelligence, does not claim to be the artificial kind, does not claim to heal or treat anything, and does not claim endorsement from people who have not signed off in writing. Where the labs win, the ledger says so. Where they lose, the ledger says that too. The three losses in the Cell Lab benchmark are on the site because hiding them would be the tell.
Where to go next
Once you know the parts, use them. Open a claim from a post or a lab page and try to walk it. If a piece is missing, that is worth telling us about.