We build a product about judgment, so it had better survive being judged.
The discipline we hold over this document and this product is the thing we're trying to teach: not being captured starts with not fooling yourself.
The goal is old — old enough that separate civilisations wrote it down independently. And for those twenty-five centuries it could only pass from teacher to student by example. There was never a reading. Plenty of people have written down the destination. Nobody has produced the evidence.
What we do is turn it, for the first time, into process evidence that is frequent, traceable and actionable. Every one of those systems has been turning, and the place they all now point at is still empty. That's where we stand.
The pedagogy people and the engineers share one table
Every design decision here — when to hand over a card, when to shut up, which sentence a report is allowed to contain — is a teaching question and an engineering question at the same time.
Two or three sentences on the judgment this person brings.
Two or three sentences on the judgment this person brings.
Two or three sentences on the judgment this person brings.
Two or three sentences on the judgment this person brings.
Two or three sentences on the judgment this person brings.
Two or three sentences on the judgment this person brings.
Holding the AI back from thinking for a child
These aren't feature limits. They're the ethics of the craft. Break any one of them and there's no reason for this product to exist.
The AI holds back, and never concludes for her
It may explain a concept, generate search terms, play the opposition, flag a broken draft. It may not set her direction, make her core judgment, or produce submittable prose. A midwife who gives birth for the mother has abolished the profession.
No manipulation
This product teaches students not to be captured. So it cannot use slot-machine mechanics to hold them — no streaks, leaderboards, badges or engagement pushes — and no coercive modals: a card may be triggered automatically, but opening it is always her call.
One question at a time
The coach is short and doesn't ramble. That saves tokens, and it protects her thinking rhythm — throw three questions at once and she'll answer none of them well.
Process is data
Skipping a card, or asking the AI to just answer, is recorded too. Friction becomes signal instead of being eliminated. A system that only measures and never intervenes just documents, with increasing accuracy, how a student became passive.
“Never for the student” means her prose and her core judgments
It does not apply to deterministic system steps: generating a plan, deriving an outline from a research question she has already stated. Those are things the system should just do. Stretching the rule to cover them makes the product harder to use without making it more principled.
Holding ourselves back from misusing a child's thinking
A product that argues for continuously observing how a student thinks has to answer a second question: on what authority, and up to where. Legitimacy isn't a default. It has to be designed.
Purpose limitation, and a list of what we won't do
The only purpose of observation is teaching and growth feedback. The axis data is never used for admissions, ranking, discipline or staff appraisal — never sold, never turned into a commercial profile. That list lives in the contract and in system permissions, not only in a privacy policy.
Minimisation and pseudonymisation
We collect only the interaction events that correspond to a measurement point; anything unrelated never enters analysis. Identity is pseudonymised in the analysis layer, behavioural detail rolls off after twelve months leaving only aggregate trajectory, and families can exercise deletion.
Tiered visibility
Different people see different granularity. The student sees her own growth report and conversation detail; parents see the cycle narrative and readiness; teachers see judgments with openable evidence; admins see class-level aggregates only. Our internal shorthand never appears in any interface.
Informed consent, and a real exit
Collection never precedes consent: informed parental consent comes before any measurement. Opting out is real — a family can use the training and stay out of the measurement, with no loss of course access and no different treatment.
High-stakes calls aren't automated, and can be contested
No judgment that could materially affect a student is made automatically. Teachers can overturn in one click, and those overturns flow back as correction data. Students and families have a channel to object, with human review inside forty-eight hours. The “insufficient evidence” state is itself an ethical design decision.
A special duty toward people who are still forming
The people being observed are children still forming. So everything is phrased as “observed this cycle”, never as a trait claim — labels have a way of coming true. And the measurement is transparent to the student: letting a child know what is worth looking at is itself part of metacognitive education.
We roll out class by class
Want to try it for your school or class? Walk the whole flow yourself first — from an unsourced post to a two-axis report — and then let's talk about rolling it out.
Structural membership (school / class) and paid entitlement are separate things: membership is a system invariant; entitlement is handled apart from it.
The school creates a class
A teacher or admin creates a class and receives a join code.
Students join with the code
Accounts bind to the right school and class. Every account belongs to a school and a class — there are no orphan accounts and no public self-serve signup.
They bring real work
EE, TOK, IA, a research project, a science-fair entry — the form doesn't matter. The thinking gets recorded either way.
Walk it yourself first. Then decide whether it belongs in your classroom.
The demo runs on real material: an unsourced post, traced all the way back to NASA and Nature Sustainability. You'll see when a card gets handed over — and what the final report says.