Gates and falsifiers

When We Were Wrong: A Running Log

Science in the open only means something if the record shows where the record moved. This page is the moving part. Every entry names the position we held, the position we now hold, and the specific piece of evidence or expert pushback that forced the change.

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. The entries below are the receipts for that stance.

How to read this log

Each entry carries a date, a "before" line (what we used to say), an "after" line (what we now say), and the trigger (what made us move). Inline claims are tagged with an evidence class: (Class E) for an expert citation or documented external pushback, and (Class C) for a configuration or integration artifact we can point at (a code diff, a copy change on the site, a schema edit, a benchmark run). Where the class is (Class U) the claim is unverified and we say so.

The log

2026-06-29, public framing
Before: Copy across the family occasionally slipped into the wrong category of phrasing, borrowing vocabulary from the AI industry rather than staying inside the natural, active-inference frame that actually describes what UNI is building toward.
After: The framing is fixed to "on the attainable path toward General Natural Intelligence, natural not artificial." A working hypothesis with growing, evidence-classed evidence, tested in the open.
Trigger: A red-team consult from the UNI Active Inference Guide GPT flagged three of five public claims as over-reaching if not constrained. (Class E) The site copy, JSON-LD, and blog cluster were rewritten in the same commit window. (Class C)
2026-06-27, review posture on the Zenodo preprint
Before: The paper was sometimes described in short-form copy as simply "the review" without status.
After: Every mention now carries the status "unrefereed preprint, expert review pending." The Zenodo record and the science page both state this. (Class C)
Trigger: A private consult with an external active-inference expert made the point that unqualified "review" implies formal refereeing to a general audience, even when the DOI page is honest. (Class E)
2026-05, Cell Lab benchmark write-up
Before: An early draft of the Cell Lab summary highlighted the wins (UNI takes 5 of 7 disturbance families against the neural baseline).
After: The published summary states the losses as plainly as the wins. Neural wins on memory_leak and cpu_noisy_neighbor. Rule-based wins on database_flaky. A single active-inference controller is not universally best, and that is precisely what the pre-registered benchmark is designed to surface. (Class C)
Trigger: An internal review of the benchmark table caught that the "wins-only" framing violated the falsifier discipline the benchmark was built to enforce. The table on the science page was rewritten to show every family and its outcome.
2026-04, autopoiesis language
Before: A working note used "autopoietic" as if it implied biological life.
After: The site fence now reads: "autopoiesis here means viable-set maintenance, not life." Free energy is the variational free energy of inference (nats), not a thermodynamic quantity. No consciousness claim. (Class C)
Trigger: An expert reader (Class E) pointed out that borrowing the vocabulary of the Maturana and Varela tradition without the fence would be a category error, and would obscure what the controller actually does.

What this log does not do

This log is not a confession booth. It is not a place to pre-empt every possible objection. It records the specific positions that moved, with the specific evidence that moved them. If you have a claim on this site you believe should be here and is not, that is exactly the pushback we want. The path is how to push back on a UNI claim, and every response we make lands in the ledger.

There is a companion piece on the ledger mechanics themselves at how we log honesty in the UNI ledger. The short version: the ledger is append-only, entries carry evidence classes, and revisions link back to the entry they superseded so the audit chain is intact.

The stance

A hypothesis is not weakened by a public revision log. It is weakened by not having one. The whole point of "science in the open" is that the moving parts are visible. Our full transparency posture, including the ledger, the falsifiers, and the honesty fences, is collected at /transparency.