HITL for School Leaders Under Pressure

You are being asked to automate a building full of learners on a timeline that was not written by anyone who has run a classroom. This piece is for the principal, the superintendent, and the department chair holding that pressure in one hand and a bell schedule in the other.

Human-in-the-loop, in the way this site uses it, is not a slogan you paste over a rollout. It is the small set of commitments that keep a school recognizable as a school after the tooling arrives. If a vendor says "the human is in the loop" and the practice inside the room does not match what follows, the phrase is doing marketing work, not pedagogy work.

What the pressure actually asks of you

The active-inference frame we work from puts this plainly (Class E, standard formulation in Parr, Pezzulo, and Friston, 2022). A mind learns by predicting what happens next, comparing that prediction to what actually happens, and repairing the internal model when the two disagree. A classroom is a room full of those loops running in parallel. Anything that severs the loop, hides the prediction, punishes the prediction error, or swaps the learner's model for a vendor's, is not a productivity gain. It is a subtraction from the very thing school is for.

Four questions to ask before you sign

You do not need a computer-science degree to hold a vendor to a real standard. You need four questions, asked in this order, at every procurement conversation and every pilot review. Our own intake, consent, and repair-loop logs on this site are configured against this exact list (Class C, this is what the workflows already require).

1. Can the learner see what the system predicted, and why?

If the platform grades, ranks, flags, or recommends without exposing its reasoning to the student, the student cannot use it to sharpen their own model. A "black-box score, with appeal" is not transparency. Transparency is the score, the features it weighed, and the version of the model that produced it, visible in the room.

2. Does a "no" from the learner count as data, or as friction?

Consent, in a classroom, is a live posture. A student who says "not that example" is signaling something the system needs to hear. If the platform logs that as noncompliance or a productivity dip, it is measuring the wrong thing. Ask the vendor to show you where a learner's refusal is stored, and what happens next.

3. When a prediction fails, does the system point at the learner or at the map?

Every learner will predict wrong. That moment is the entire opportunity. A repair posture treats the failed prediction as information about the model that needs updating, and gives the learner a small, testable next move. A correction posture treats it as a defect in the learner, marks it, and moves on. Ask the vendor how their system frames a wrong answer on screen. That language is your policy, whether you wrote it or not.

4. Where do the teacher's judgment and the family's context enter the loop?

A school is not one loop. It is a nested set of them: learner, teacher, family, cohort, building, district. If the platform closes the loop between learner and vendor server and treats the teacher as a monitor rather than a co-modeler, you have adopted a different institution and not told anyone.

What this posture rules out, plainly

A tool that grades a student and offers no path to inspect the grade is not in the loop with them. A behavior-flagging system that produces a risk score the family cannot see and cannot contest is not a safety measure. A vendor that treats teacher overrides as error rates to be minimized has confused its metric with your mission. None of that is human-in-the-loop. It is human-under-the-loop with a friendlier logo.

Where our own work is not yet sure, we say so. Whether the four-question filter above holds up across many districts and vendor categories over a full school year is (Class U) not yet demonstrated at that scale. We are running it with the leaders who choose to work with us, and we will publish the falsifier when we have enough runs.

One outside signal worth reading

Zooming out from the building: the wider job-market conversation your graduates are walking into is being shaped in public. AJ Maren's Meet Jay Kumar Chimata: JobFirst.ai and the Real AI Job Market is a useful outside view of that terrain, and in our voice it is a signal, not an endorsement, that the automation story your students hear at home is more contested than the marketing suggests.

A quiet definition, for a hard job

Human-in-the-loop for a school leader means this. The learner holds the model. The teacher and the family sit at the table with the learner, not above them. The platform offers guesses, evidence, and structured surprise, on the record, with the reasoning visible. A refusal counts. A failure is read as an invitation to repair the map, not as a verdict on the person. Your signature on a contract is what protects that arrangement, or fails to. Ask the four questions in the room, and the room stays a school.