Everything in Mathiter, and the state each part is actually in.

Five to thirty questions, each chosen from the answers before it. It stops when the estimate is precise enough rather than when a question count runs out, so a student who is consistent finishes early. Free, uncapped, and it stays free — a student can sit it again next term without anything changing on your side.

A two-parameter IRT model, estimated expected-a-posteriori over 61 quadrature points, stopping at a standard error of 0.35 or 0.45 depending on the run. What comes back is an ability figure, the range it is confident within, and a percentile. Not a grade equivalent, and not a score out of anything.

The student's track laid out as it is actually taught, marking what is finished, what is in progress and what comes next.

Placement and practice data are analysed into a projected Digital SAT score. That is the only exam that returns one. Six others have a configured profile in the code and none of them produces a number — a request for any of them fails rather than guessing, which is the behaviour you want but not what a feature table would have told you.

The pool has no questions below Grade 5, so the test cannot place a primary student. What it would return for one is a floor, not a measurement — so it is not offered.

Item difficulty is calibrated across the whole population. There is no per-school, per-class or per-cohort scope anywhere in the calibrator. Your students' answers improve the pool for everyone using it, including schools that are not you, and the estimate your students get is not sharpened by being yours.

Every wrong option in every question was written on purpose, and carries the name of the misconception that produces it, the student reasoning behind it, and an ordered chain of steps that walks back out — with a hint at each step and the correct answer withheld to the end. None of it is generated when the student clicks. It was written beforehand and stored, which is why the same wrong option teaches the same thing every time, and why you can read the teaching before a student ever meets it.

A wrong idea a student has met becomes a card that can be put in front of them again, sorted into ten kinds — among them sign errors, order of operations, misreading the question, skipped steps, units, and confused variables. Each card records how often it has come back and whether it has been overcome.

By topic, weakness-first, mixed, and timed.

Wrong answers collect for re-solving without anyone tagging them, alongside bookmarked questions and the full attempt history.

A student types a question, pastes or uploads an image up to 5 MB, or attaches a PDF up to 10 MB, and gets a worked explanation back. Ten a day on a free account, a hundred on a paid one. The questions and answers stay in the student's review history.

The history belongs to the student's account. A teacher has no view of what a student asked, so this is not a channel you can supervise.

Accuracy, problems attempted, streaks and activity across the students assigned to you, and the class's accuracy on each individual question — which is the view that tells you a question was the problem rather than the students.

Everyone under 60% accuracy, with their accuracy and problem count. It does not tell you the topic: the screen has a slot for one and the server does not send it, so that line renders empty. It is the most useful thing this list could say and it is the thing it does not say.

For one student at a time: the ability estimate, a radar across strands, and which misconceptions they are carrying and how often.

Build practice from the curriculum tree and assign it. The tree is hardcoded to the US system — Grade 6 through AP Calculus AB costs you nothing, and IB or Cambridge costs you the builder. The standards and the questions exist for those tracks; the builder cannot walk to them.

One teacher and one student in a shared session. It is not a class and calling it one would be a lie of scale.