A year is long enough for a school to notice a pattern, and short enough that we should not overclaim one. This is what we saw, and what we still do not know.
The setup, in one paragraph
We spent an academic year with one partner school. One cluster of teachers, one department head, one building leader. We used the vocabulary of active inference (generative models, prediction error, prior updating) inside existing curriculum, not as a replacement for it. Our integration was scoped to lesson design, rubric language, and a small set of shared classroom routines. It was a configuration effort more than a technology deployment (Class C).
What changed, said carefully
Teachers reported that rubric language became more precise. When a student wrote something confident and wrong, the rubric now had a place to say so without punishing the confidence itself. Prior, error, revision. That vocabulary carried. (Class E: teacher self-report across the cluster, gathered in end-of-year debriefs. Not a controlled study.)
The department head reported fewer arguments about grading edge cases. Not zero. Fewer. The shared vocabulary gave staff a way to disagree about a specific piece of student work without disagreeing about pedagogy at large. (Class E.)
We saw the AI-authorship fence hold. Every student artifact that involved outside tooling carried a note about what tool did what. That was a configuration change to the assignment templates, and it stuck because the templates stuck (Class C).
What did not change
Test scores. We did not see a signal we would stand behind in either direction over one year, with one cluster, without a comparison group. Anyone who tells you a single-site, single-year integration moved standardized outcomes is selling something. We are not.
Workload. Teachers who adopted the vocabulary spent about the same total planning time. The time shifted, from rubric arguments toward rubric design, but it did not shrink. We consider that honest, not a failure.
Parent understanding. Most parents did not read the vocabulary note we sent home. The ones who did asked good questions. We are still learning how to make that note land for a busy household. If we claimed we had figured out parent communication, we would be lying.
What surprised us
Two teachers who were skeptical at the start became the strongest internal advocates by spring. Their skepticism was useful. They pressed on every claim we made, and the ones that survived their pressure are the ones we now trust more. We would rather have two rigorous skeptics than ten enthusiastic adopters, and we would say that to any school considering a pilot.
What we would do differently
Start with the rubric, not the framework. The vocabulary landed when it entered rubrics. It stalled when it entered slide decks. If we did year one over, we would spend the first month rewriting one rubric together, and skip the framing session entirely.
Publish more of the mess. We kept internal notes on what did not work and shared them only with the partner school. Some of that should have been public. Practitioners learn from other practitioners' failures faster than from anyone's successes.
The honest posture
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.
This post is one data point (Class E), gathered inside one configuration (Class C), across one school year. It is not a study. It is a reflection. Weight it accordingly.
Where to go next
- If you want to see the guardrails we hold ourselves to inside a partnership, read what SolutionWright does and does not do.
- If you want the measurement approach we used this year, read how we measure a pilot honestly.
- If you are wondering about the exit door, read when UNI fails, what we do.
- If you are ready to talk about a scoped year of work at your school, the /workshop page has the shape of it.