This document is about why that measurement can be trusted, what follows if it can, and the date on which that is decided.
Why education has never had a verifier
AI improved fastest where it could check its own work. Code has tests and mathematics has proofs; the loop between attempt and verdict closes in seconds and runs millions of times. Science and medicine move slowly because their verifier is the world, and the world takes months to answer. Education sits on the slow side with a worse problem. Its fast measures, exam scores and rubrics applied to prose, are destroyed by the act of optimizing against them. This is one of the best-documented facts in the field: scores on high-stakes tests rise for years while scores on low-stakes tests of the same material do not move (Campbell named the pattern in 1976; Koretz has documented score inflation across decades of testing programmes). The students learned the test. A model tutor steered by an exam rubric is that failure made cheap and fast. It will produce fluent test-takers without deep understanding of the subject.
Everyone building in education with models can see this, and there have been two responses. Neither produces a verifier.
The first keeps the old objects, the lesson, the question, the essay, the rubric, and puts a model on top. Those objects were shaped by scarcity: humans had to author, humans had to score, classrooms had limited hands. The constraints are gone, but the proxies they produced are still the thing being optimized, so the loop closes against the exam.
The second sees that the standard hollows out the disciplines and answers by redefining what should be learned: critical thinking, reasoning games, a curriculum of the founder's own design. This is the more serious response and it has the deeper flaw. The target has moved from the discipline to the founder. Nobody outside the company can check it, so nobody can find out it is wrong, and a target that cannot lose cannot be iterated against. A founder's account of what reasoning is ends up competing with physics, history and mathematics, with no way to be corrected.
A verifier needs a target that sits outside the measurer, can be checked by someone else, and can lose. The disciplines already have one.
The inversion
Nobody worries about teaching to the test in chess, in code, or on a running track. The board, the compiler and the stopwatch cannot be talked into anything. In those fields the test is the activity itself, so getting better at the test and getting better at the thing are one process, and the loop can run as hard as anyone likes. Education's tests are not like that. They stand in for the thing, and the section above is the story of what happens to a stand-in under pressure. The question Norem asks is whether a test of reasoning can be built that behaves like the board: one that nothing can move except the reasoning.
History has a board too. You can sound like a historian without being one, and you can pass the history exams without being one. What you cannot do is write a historical argument that survives review by historians. Norem's test is that review, taken one move at a time.
The learner commits before the evidence is shown, and each commitment is a prediction about a document they have not yet read, resolved by that document afterwards. A prediction is the one judgment fluency cannot fake: a grader reads reasoning off the surface of prose after the fact, and the surface can be learned, while a forecast about an unread document cannot be. The learner's prose is kept and logged; it is not what is scored.
Reviewed against whom, is the other half. A model grading prose reviews against itself; a founder's curriculum, against the founder. In history the reviewers' judgment already exists outside both: every work of history is an expert's judgment on the documents, defended in public, sitting on a shelf. This is the move everything turns on. An expert's judgment has never scaled because it had to be given live, one student at a time. This one is already written. Norem shapes it so that a machine can score every learner's work against it. The one thing left for a human is to confirm that the shaping is faithful to the authors; a working historian signs that it is. The machine is not grading on its own opinion; it is applying a signed one. The key sits outside the company and can be shown wrong by any other historian, which is what a target has to be.
A test that can only be passed by doing the discipline's reasoning, scored against the discipline's own judgments, can be taught to without harm, because teaching to it is teaching the discipline. The failure that has defeated every measure education has had becomes the mechanism. That is the inversion, and everything Norem is rests on it.
What it unlocks
If the measure holds, education becomes the next thing AI gets very good at: not explanation, which it already has, but teaching that leaves a person able to do what the discipline does, and that gets better every time it is used because it can tell when it has worked. That has been imagined for fifty years and delivered by nobody. No curriculum, no lecture. A learner sits down and is met by the one encounter that is right for them at that moment: the question that will find what they have misunderstood, the exercise that forms around the misunderstanding, the challenge that is just slightly too much for them, which is the only kind that moves anyone. They do the discipline rather than read about it, and what they do decides what comes next. Every generation of educational technology has promised this. None could deliver it, because none could tell whether the last encounter had worked, and so none could choose the next. That is the whole of what a verifier is for.
It is a company rather than a paper because of what cannot be copied. A model can reproduce the method, the interface and the reading of a historian's book within weeks, and Norem assumes it will. What it cannot reproduce is the slow part, for the same reason it can design a drug trial and cannot run one: a historian's signature on each body of material, learners met under research conditions, and a public record of claims filed before their data. Those are the inputs. They have always been expensive; what changed is that everything around them became nearly free, so for the first time a company can be organized around them alone. That is the whole of why now. What compounds is what they produce. Every instrument built and every cohort run teaches Norem what to build next, so each instrument is cheaper and its evidence better than the last. Money can buy the inputs. It cannot buy the turns of the loop.
History is the first instrument, not the company. The prototype was built there, on the founder's own historiographical work, and it is a good first test because doing the act and writing about it come apart visibly in history, which is where a grader of prose is weakest and a measure of commitments has the most to show. A measure that holds there could hold anywhere a discipline's judgment is written down: medicine, law, audit. Nothing outside history has been tested. It scores a model as readily as a learner. The ceiling is every field where someone has to reason, and someone else has to be able to tell. The first buyers sought are examination-preparation providers and school programmes with essay-based history assessment, whose question is whether a student did the source work or wrote the template. What Norem sells beyond that, tutors, evidence or experiments, is decided by what buyers pay for, under a rule written before the first commitment comes in.
What it does not claim
Every measure is a proxy, this one included. This one measures the moves, not the mind: it can tell whether someone can do what historians do every day, not whether they would make a good historian. Where the discipline has no settled answer, the test does not invent one; more than half of what was tried on the first material was turned away for that reason. And a test can be impossible to fake and still not matter. Only what the learner can do afterwards, on something they have never seen, can show that it does.
Fields become sciences when their experiments become cheap enough to run many of them. Education has had the right target before: history-education researchers have mapped the components of historical reasoning in detail, and the intelligent-tutoring lineage in mathematics measured those components per step and improved teaching against them. That lineage worked where it was tried and stalled on authoring cost: hundreds of expert hours for each hour of instruction. What has not existed is a measure of that resolution that is cheap enough to run many experiments against and that survives being optimized against. That is why Norem is built as an apparatus for running experiments rather than a product with an opinion. The first three are below.
The tests, and the date
This document has made three claims: that a measure of reasoning can be built which nothing moves except the reasoning; that it holds when a tutor is steered by it; and that what it unlocks is worth a company. The third cannot be tested yet, and February will not touch it. The second is tested when a cohort meets for more than one session. The first is tested now. Each is filed on a public record before its data exists, so the result cannot be written afterwards, and we are seeking an outside co-signer for each.
The first experiment: does the measure beat a frontier model grading the same learners' written arguments? Blind readers decide every disagreement. It holds if they side with the measure two times in three. This tests the grader, not the teaching.
The second: can it be fooled? Answers written to score well without doing the reasoning, by models and by people paid to try, go to both graders. It holds if the measure is fooled at most half as often as the model, and at most one time in five, without rejecting more honest work.
The third: does a tutor steered by this measure teach better than one steered by a model's grading, on something neither grader sees? It needs learners who come back, so it is filed later. It is the only one that can say whether the measure matters. If it fails, the environment above is not built on this measure, whatever the first two showed.
If the first two both fail, Norem says so and reconsiders within two weeks: another domain, or stopping. If one fails, it is named in every claim Norem makes. The pass marks are fixed when the entries are filed on 20 November 2026; the filed versions count. About fifteen learners take a ninety-minute session in January 2027 on material they have never seen. On 19 February 2027 both results go on the record, whichever way they go. That is the date.