Something we say in class does not survive contact with a student's question. The gate we built to check understanding lets the wrong answer through. A public paragraph on our own site turns out to overreach. What happens next is the whole point.
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 posture is only real if we have a practice for the moments it breaks.
Three shapes of failure
We treat three failure shapes as normal, not shameful.
A gate fails in class. A gate is a small, teachable check: predict, then check against a source, then update the prediction. When a class of students routes around the gate, or the gate accepts a fluent-sounding wrong answer, the gate is broken. We record the gate id, the class of failure, and the fix, in the same document the teacher reads before the next lesson (Class C).
A claim wobbles under a student's question. A student asks the sharp form of the question we did not want asked. We pause. We say the honest version out loud: "That is a good question, and my previous sentence was stronger than the evidence supports." Then we restate the claim at the altitude the evidence actually reaches, and mark it with an evidence class (A empirical-in-session, B code and inspection, C configuration and integration, E expert citation, F falsifier present, U unverified).
A public statement needs to be retracted. Something on the site, in a slide, or in a workshop handout crossed a red line: we implied endorsement we do not have, we used a banned word for our own work, or we described a result more strongly than the receipt supports. We change the page, note the change in a visible retraction log, and, when the reach warrants it, we say what we changed and why. The change is the record.
The rule set we teach students
We teach students the rule set we use on ourselves. There is a working canon we cite: Parr, Pezzulo, and Friston (2022) on active inference (Class E). There is a public record of what we build: source, gates, and configuration on the transparency page (Class C). There are words we do not use for our own work, because they claim more than we can show. And there is a rhythm: predict, check, update, and, when we are wrong, say so at the same volume we said the original claim.
Students learn that "I was wrong, here is the corrected version" is a stronger sentence than "I was right all along." A retraction is not a loss of face. It is the receipt that the method is working.
What we do not do
We do not quietly edit a paragraph and pretend the earlier version never existed. We do not blame the student who found the crack. We do not fold a broken gate back into the syllabus without a note. We do not describe a hypothesis as if it were a finished result. And we do not use the words a marketing team would reach for when the evidence is thinner than the sentence.
The classroom cadence
Weekly, we look at three columns: gates that passed, gates that failed, and claims we softened. The point is not the tally. The point is that the columns exist, in the open, and that students see the teacher writing in the middle column without flinching. That is what a working hypothesis looks like when it is not a slogan.
Where to go next
- Working hypothesis as a classroom stance: the posture this practice grows out of.
- What UNI is and what it is not: the perimeter we hold for our own work.
- Transparency page: the retraction log, the gate registry, and the receipts.
- The workshop: where teachers and school leaders practice the cadence with us for one full day.