How it works
What MathForge measures, how it measures it, and what its numbers do not mean. Written for students, and for anyone who wants to check the reasoning.
1. Where the questions come from
Every question is original. Each of the 118 question types is a small program that, given a seed, picks numbers so that the answer comes out clean, writes the question, and writes the worked solution alongside it. The same four values — question type, seed, level and format — always rebuild exactly the same question, which is how your notebook can store a mistake in a few bytes.
Wrong choices are not random numbers. Each one is the result of a specific, named mistake — forgetting to flip an inequality, dividing by the new price instead of the original — and that mistake is what you are shown if you pick it.
Answer keys are checked independently. 104 question types carry a machine-checkable statement of why the answer is right. In the test suite, a math engine written from scratch for this site evaluates that statement for 30 seeds of every variant — by direct evaluation, numerical differentiation, numerical integration or a two-sided limit — and compares it with the key. The same run compiles every formula and rejects any question without four distinct choices. A wrong key cannot be merged.
2. The response model
A question is treated as a three-parameter logistic item. The probability that a student of ability answers it correctly is
where is the difficulty (set by the level: foundational, standard or challenging), is how sharply the question separates students, and is the chance of guessing it: for four choices, 0 for a typed answer. Ability is estimated by Bayes’ rule on a grid of 81 values from −4 to 4, so every number on the site is an exact posterior calculation rather than an approximation.
3. Mastery, and why a few answers are not enough
Ability is modeled at three nested levels — the course, its domains, and the 36 skills inside them:
The consequence is credibility, in the sense actuaries use the word. If you have answered two questions on circles and got both wrong, the site does not conclude that you cannot do circles. It starts from how you do on the rest of geometry and moves away from that only as evidence about circles accumulates. With thirty answers, the thirty answers dominate. The posterior is computed exactly, by passing messages up and down the three-level tree.
Recent answers count more. An answer’s weight halves every 30 days, so the report follows what you can do now.
Uncertainty is shown, not hidden. The percentage is the estimated probability of answering a standard-level question on that skill correctly without guessing. The lighter band on each bar is one posterior standard deviation either side. Below 1.5answers’ worth of evidence, a skill is labeled “Not enough data” and given no number at all.
4. How the next question is chosen
Practice uses Thompson sampling. For every skill, the site draws one plausible ability from that skill’s posterior, and practices the skill whose draw came out lowest. Weak skills are chosen most often; skills the site is unsure about are still visited, because a wide posterior sometimes produces a low draw. The level is then matched to your estimate, aiming for a question you should get right about two times in three.
Missed questions enter a Leitner notebook and return the same day, then after 1, 3 and 7 days; four correct answers in a row clear a card. A card is reviewed with a fresh variant — the same idea with new numbers — so it cannot be passed by remembering an answer.
5. Mock tests and the score estimate
A full mock test follows the published structure of the digital SAT Math section: two modules of 22 questions and 35minutes each, with roughly a quarter of the questions answered by typing, and the four content areas in the section’s proportions (7 Algebra, 7 Advanced Math, 4 Problem-Solving and Data Analysis, 4 Geometry and Trigonometry per module).
After the first module your ability is estimated from those answers alone. At or above you are routed to a second module weighted toward challenging questions; below it, toward foundational ones. The final estimate uses all of your answers and is mapped linearly onto the 200–800 scale, rounded to the nearest 10.
What these numbers do not mean
The item parameters are set by design, not fitted to data. A testing organization calibrates each question on thousands of students. This site has no server and collects no responses, so it cannot. Difficulty comes from the level a question was written at.
The score is an unofficial estimate. It has never been equated against real test results. It is useful for comparing your own tests over time; it is not a prediction of an official score, and it is always shown with a range for that reason.
Coverage is broad, not complete. The bank spans the main skills of each course, but a real exam can ask things that are not here.
Your data stays with you.There are no accounts and no analytics on your answers. Progress lives in this browser’s storage and can be exported from the Progress page.
The full source, including the tests described above, is on GitHub. SAT® and AP® are trademarks registered by the College Board, which is not affiliated with, and does not endorse, this site.