PrepGraph Research
What a student knows should follow her.
We study synthetic memory, and we write for principals and teachers first.
- What we study
- PrepGraph Research studies synthetic memory: an AI-held, learner-owned account of what a student knows and can do.
- From the classroom to her first job
- It is checked by her teachers and carried to her next school and her first employer.
- Who we are
- We are the research group of PrepGraph, which works in India and the United States, from the classroom to the first job.
Abstract
Consider one Tuesday. One teacher, fifty-five students, forty minutes, and a syllabus with a date on it. The only instrument she is held to is the mark, and the mark arrives after the unit has closed. A common paper, set and marked outside the school, is among the fairer instruments Indian schooling has: comparable across schools that share nothing else, and hard for anyone inside a school to curate. It is a serviceable instrument for ranking children and a poor one for teaching them, and Indian schooling uses it for both, because at the moment a decision about a child is made it has nothing else it trusts.
What the mark discards is never general. It is always the particular thing the teacher would have needed: which topic, which misunderstanding, which idea from an earlier year that everything since has rested on, which skill, and worst of all which moment. The week in which a small error was ten minutes of work opens early in a chapter and closes long before the report card is printed. By March the wrong method has been practised until it is the child’s settled method.
The obvious repair is to describe the child more richly, and that is harder than it sounds. A wrong answer is ambiguous: the idea may never have been understood, the question may have been misread in a language not spoken at home, the arithmetic may have slipped, or the clock may have run out. A classroom produces thin evidence about any one child, work is shared and copied, and the teacher has forty minutes. A description of a child is also easier to curate than a common paper marked by a stranger, so a finer account offers more surfaces to perform on, not fewer.
This group studies one form that richer description could take: synthetic memory, an account of what a learner has shown that an AI tutor keeps for her, that her teachers can check and contradict, and that goes with her when she changes teacher or school and, later, when she leaves school for further study or work. PrepGraph’s tutor already keeps a memory a learner can read and erase. How such an account is kept true, carried and checked is the research.
So this group publishes its questions first, and states with each what would tell us to stop. The first question below is abandoned if a claim about one child changes nothing about what her teacher does that week, or if two teachers reading the same student work cannot be brought to name the same thing. Every study we run will be published with its method, so that anyone without our interest in the outcome can check it.
Three students in the same class sit the same mathematics paper. All three score sixty-two out of a hundred. All three are entered in the register as average, and if the school runs an extra class, all three are sent to the same one and given more of the same paper.
She can carry out every procedure the paper asks for. In each long question she loses marks to one habit: when she distributes a subtraction across a bracket, she drops the sign.
NeedsTen minutes on a single habit, this week, before the wrong method becomes the method she trusts.
He answered the quadratics section without a mistake and could not begin the word problems. He cannot turn four lines of English into an equation, which is not a difficulty in this year’s chapter and will stop him again in physics.
NeedsWork that moves between a sentence and an equation, because more mathematics practice will not touch the thing that stopped him.
She knew the material and answered the last two questions in note form because the clock ran out. Reading each question took her much longer than it took the boy beside her, in a language she does not speak at home.
NeedsA response to reading under time rather than to mathematics: recorded as a subject weakness, it will be remediated in the subject and therefore not remediated at all.
One number, three children, three different things to do. Two of them do not need more mathematics, and the extra class will hand all three the same paper again. Nothing in the register separates them, so nothing in the timetable does either.
Definition
Synthetic memory
An account of what one learner has shown, kept for her by an AI tutor, so that her teachers can act on it this week and she can take it with her when she moves on.
Synthetic, because the account is made: an AI tutor builds it from what the learner has actually done, where today the only lasting account of a child is the one a teacher carries in her head. Memory, because it lasts: it outlives the lesson, the term and the teacher who formed it, which is the thing a teacher’s own knowledge of a child cannot do. The distinction that matters inside it is between a record and a model. A record states what a child produced: this answer, that score, this many minutes, on that date. Schools already hold far more records than anybody has time to read, and moving a record onto a screen does not change what kind of thing it is.
A model states what a record cannot: what a child knows, what she can do and under which conditions, and what she is ready for next. Those are three different claims about a person, not one claim with three names. Knows is a claim about a present state nobody can observe directly. Can do is a claim about a capability under conditions, and the conditions do more work than anyone admits. Is ready for is a claim about a future, inferred from a past that was thinly observed. Anything that runs the three together will be confidently wrong in a way that is hard to notice from outside and harder from inside. On this page the term names the object of the research, and anyone who uses it, ourselves first, can fairly be asked which of the three claims they mean and what would show them to be wrong about it.
What it is not
- Built with AI and answerable to people: the account is kept by an AI tutor, and it is meant to be checked, contradicted and corrected by the teachers who know the learner.
- Not a log of activity. Minutes, clicks and attempts are records of the application; they are not knowledge of the learner.
- Not a measure of how clever a child is. It concerns what she has learnt and what she needs next, which are different questions.
- Not a richer score. Replacing one number with a slightly better number inherits every problem described above.
Open questions
Five questions, each with the finding that would end it
Broader questions about learning are left to the people already working on them. These five are the ones this group has chosen to answer, because each names something that breaks in an ordinary section of fifty-five.
- 01 One child, one lesson
For one named child in one named class, can what she needs next be said in words her teacher can teach on Tuesday, and put in a form that lets the teacher find us wrong by Friday?
A teacher with five sections cannot afford a diagnosis that takes longer than the remedy.
Stops ifWe stop if a claim of this kind, put in front of the teacher who could act on it, changes nothing about what she does that week — or if two teachers reading the same student work cannot be brought to name the same thing.
- 02 Different, or merely easier
If two children in one class are given different work, what has to differ about it, beyond how hard and how much, for each of them to be better off a fortnight later?
The familiar failure gives a struggling child easier and easier work, so that she answers correctly while practising around the missing idea instead of through it.
Stops ifWe stop if the only difference that survives scrutiny is quantity: work that differs only in how much and how hard is more practice, and should be called that.
- 03 Something a school can watch
What can a school observe about one section over one term that moves when learning moves and stays still when only activity moves, at what interval, and at what cost in teacher time?
Counts of activity rise the moment somebody is told to raise them, and the mark at the other end arrives too late to decide anything.
Stops ifWe stop if the observation can be raised by telling teachers to raise it with nothing changing in what students know, or if a school that makes it decides nothing differently for having made it.
- 04 What the teacher keeps
Which teaching tasks can be taken off a teacher without taking away the knowledge of her class that performing them gave her, and which must never leave her at all?
Marking is how a teacher finds out that eleven children made the same mistake; hand it over and she receives a report while the thing the report was meant to carry has gone.
Stops ifWe stop if teachers whose marking is automated become less able to say who in their class is lost, and that knowledge does not return by some other route. The design would then be wrong however much time it saves.
- 05 What survives July
When a child changes teacher, section or school, what should travel with her besides a mark sheet, who should own it, what should she be able to erase, and what does it cost her when it turns out to be wrong about her?
A rebalanced section, a transfer certificate or one resignation erases a working model of a child that took a year to build, and she arrives at the next desk as a stranger with a number attached.
Stops ifWe stop treating this as a question about learning if transitions go no better with carried context than without. It would remain a question about a person’s rights over what is held about them, which is a different argument and would have to be made on those grounds.
Intentions
What this is meant to put in a school’s hands
These are the things the work is built to put in a school’s hands. We state them here so that every school can hold us to them.
- A way of stating what one child needs next that is precise enough for her teacher to contradict, published together with the times teachers contradicted it and were right.
- Something a school can observe about its own sections during a term, described plainly enough to run without us, to copy without us, and to read against us.
- A list of teaching tasks we decline to automate, with the reasons, that does not shrink when it becomes commercially inconvenient.
- An account of a child that belongs to the child rather than to the school that generated it or the company that holds it, and that can be taken elsewhere, including to a competitor.
- The status of each research direction — sustained, revised or retired — with the date it changed, so that anyone can follow the programme as it moves.
How we work
How the research is run
Every question on this site is published with the finding that would end it, and every study we run will be published with its method.
- Where it starts
- PrepGraph’s tutor, at prepgraph.com, keeps a memory of what each learner has shown, and the learner can read what it holds and erase it.
- The agenda
- Five questions and four research directions, each published with the finding that would end it.
The practice every study follows
- Stated in advance
- Each question and direction is published with its stopping condition before any study of it begins, so the condition cannot be written to fit a result.
- Before a study starts
- Its question, method and stopping condition are written down, and anything involving students is cleared by an ethics review and agreed with the school in writing first.
- When a result arrives
- The study is published with its method, its setting and its sample, so it can be checked rather than trusted, and the direction’s status is recorded with its date.
PrepGraph is an AI learning company, working in India and the United States, from school to first job. It sells learning software to students, families and schools, and runs a research programme on how learners remember and build skills. Every finding of ours that bears on PrepGraph’s own products will be published with its method, so that anyone without our interest in the outcome can test it.
Published
- Positions, argued, that a reader can disagree with
- Research directions, and the condition that would end each
- Open problems, stated precisely
- Every study we run, with its method attached
Kept private
- Methods, architectures, parameters
- Models, vendors, systems
- Platform figures
Three kinds of work
Research, responsibility, and what ships
Research discovers what is possible. The Council determines what is responsible. The products demonstrate what can be implemented.
PrepGraph Research
Research
Research on synthetic memory: questions, positions and principles. This site.
PrepGraph
Corporate
The company’s own site, from the classroom to the first job: what PrepGraph is, the tutor students and families start with, and the prices of everything it sells.
PrepGraph Council for Responsible AI & School Innovation
Non-commercial forum
A forum on responsible AI in Indian schools, convened and funded by PrepGraph. It certifies nothing and endorses no vendor.
PrepGraph Academy
Professional development
Professional development for teachers and school leaders.
PrepGraph Courses
Product
Where PrepGraph’s courses and certifications are found and bought by individuals.
PrepGraph Talent
Product
PrepGraph’s hiring marketplace: a skills passport, work simulations and introductions to employers, only with the candidate’s consent.
PrepGraph for Institutions
Product
The School Programme and partnerships for institutions, and the portal schools use to run their side of the platform.
PrepGraph for Charter Schools
Product
Where PrepGraph offers the platform to US public charter schools, with a pilot quoted in writing.
altme.org
Separate initiative, not a PrepGraph property
A separate initiative on a person’s rights over their own memory.
Correspondence
If you run a section of fifty-five, we want to hear about your week: the question that matters most in your classroom, the one we have put in a way no teacher would recognise, or what you needed on the Tuesday the mark could not tell you. There is nothing here to buy and nothing to sign up for. Write to [email protected].