A student can now produce a fluent paragraph on almost any topic in under ten seconds. The rubric that rewards fluent paragraphs is, as of this school year, measuring the wrong thing.
This post is for the curriculum designer who needs a rubric that survives that shift. It walks through gate-oriented assessment: observable checkpoints where a student has to predict, revise, and cite, in that order. It is grounded in active inference (prediction, error, update), and it treats "evidence classes" as a classroom vocabulary rather than a research footnote.
Why fluency-of-output rubrics fail now
A rubric that scores "clear thesis, three supporting points, conclusion" is scoring the shape of a paragraph. That shape is free (Class B, direct inspection of any current chat model output). The signal that used to sit inside the shape (a student having actually thought the thought) has separated from the shape.
The teaching response is not to ban the tool. It is to move the rubric to a place the tool cannot ghost-write for the student in real time: the moment of prediction before evidence arrives, the moment of revision after evidence arrives, and the specific citation trail that connects the two. Those three moments are what a gate observes.
What a gate is
A gate is a checkpoint in a lesson that a student either passes or does not, based on an observable act, not a produced artifact. "Observable" means the teacher (or a peer, or the student themselves against a key) can point to what happened.
Three gate types cover most curricula:
- Prediction gate. Before the reading, before the experiment, before the source is opened: the student writes a specific, falsifiable prediction. "I think the character will refuse" is a prediction. "I think it will be interesting" is not.
- Revision gate. After the evidence arrives: the student writes what changed in their prediction and what specifically caused the change. If nothing changed, they say why the evidence was consistent with the prediction. A revision that ignores the evidence fails the gate.
- Citation gate. The student points to the exact sentence, paragraph, data row, or timestamp that drove the revision. A vague "the article said" fails the gate. A quoted line with a page or paragraph number passes.
Notice what a gate does not ask for: polish, length, style, or a final answer. A student who ends the lesson with a wrong conclusion but a clean prediction, revision, and citation trail has passed the gates. A student who ends with the right conclusion but no trail has not. That inversion is the whole point.
Why this maps to active inference
Active inference frames a mind as a system that maintains a generative model of its world, makes predictions, notices prediction error, and updates the model to reduce future error (Class E, Parr, Pezzulo, and Friston, 2022, Active Inference: The Free Energy Principle in Mind, Brain, and Behavior). Prediction, error, update.
A prediction gate makes the student's generative model visible before evidence contaminates it. A revision gate makes the prediction error and the update visible. A citation gate makes the evidence that drove the update auditable. The three gates are, in classroom form, the loop the theory says learning already is.
This is not a claim that active inference has been established as the settled mechanism of human learning. It is a claim that structuring assessment around prediction, error, and update is consistent with a serious, published research program (Class E), and that it produces observable classroom artifacts a teacher can grade (Class B). The framing is a working hypothesis, tested in the open, evidence classed as it accumulates. That posture, science in the open, is the same one we use across the family of sites.
Evidence classes as classroom vocabulary
The evidence-class tags we use internally (A empirical in session, B code or direct inspection, C configuration or integration, E expert citation, F falsifier present, U unverified) transfer directly to student work. A middle school student can be taught to write, next to any claim in their revision:
- (A) I saw this happen in the experiment today.
- (B) I checked the source text and this is what it literally says.
- (C) This depends on the setup we agreed on at the start.
- (E) A published expert says this, and here is the citation.
- (F) Here is what would make this claim wrong.
- (U) I have not verified this yet.
A student who tags their own sentences this way is doing epistemology as a habit. A student who writes an "F" line, a specific way their own claim could be wrong, is doing something a fluent LLM paragraph almost never does on its own. The absence of an F line is itself a gate signal.
How to score falsifiability
Falsifiability is the gate that raises the ceiling of the whole rubric. A claim without an F line is graded as unverified (U) by default. A claim with an F line that is trivially untestable ("unless the sky turns green") fails the gate. A claim with an F line that specifies an observation the student could actually make ("if we run the experiment again with cold water instead of warm and the result is the same, my mechanism is wrong") passes.
Scoring is binary at the gate level (pass, fail) and cumulative at the lesson level (how many gates passed out of how many attempted). The reason to keep gate scoring binary is that it removes the room for a graded rubric to reward polish. Either the student wrote a falsifier a peer can evaluate, or they did not.
What this looks like on a unit plan
A unit built this way has three visible columns per lesson: what the student predicted, what they revised to, and what evidence they cited. The teacher's planning question shifts from "what will they produce" to "what will they have to predict before evidence, and what evidence will force a revision." That is a different kind of lesson plan, and it is the one this site is built to support.
A related post walks through the concrete workbench a teacher can keep on their desk to run this in class: see the teacher's workbench tour. The vocabulary side (how to introduce A, B, C, E, F, U to students without it feeling like jargon) is covered in evidence classes as classroom vocabulary. A worked middle school example is in gate design for middle school. And the framing question of why the target of instruction is the underlying thinking, not the LLM tool, is in teach UNI, not LLM tooling.
Where this fits in the wider map
A useful public map of where active inference work is happening in 2026, and where this site sits among adjacent efforts, is here: Where to Start with Active Inference, a Resource Map for 2026. Our one-line frame, in our voice, not paraphrased from theirs: it lists SWU among five pathways into the field, which is how a curriculum designer new to active inference can orient before choosing a route in.
Next steps
- Bring a unit you already teach to the /workshop and rebuild one lesson around three gates, prediction, revision, citation, before you rewrite anything else.
- Read teach UNI, not LLM tooling for the framing that keeps the target of instruction on the student's thinking.
- Walk through the teacher's workbench tour to see the daily artifacts this rubric produces.
- If you want a worked example before designing your own, start with gate design for middle school.
AI-authorship fence: this post was drafted with LLM assistance, reviewed and edited by a human author, and is published under human editorial responsibility. No claim is made that an LLM authored the reasoning; the framing is a working hypothesis on the attainable path toward General Natural Intelligence, natural not artificial, tested in the open.