How Preptics works

• Our software does the maths. Every score, range and estimate is calculated by code we can explain. When AI writes about your results, it may not state a number of its own.
• Ranges, not invented precision. An estimate is shown with how uncertain it is.
• Estimates never change a score. They decide what you practise, not what you scored.
• No AI inside a timed test. The only help in a sitting is what the real test gives you: the calculator and the reference sheet.

Reading and Writing is two 32-minute modules and Math is two 35-minute modules. In each section the second module is easier or harder depending on how you did in the first. Every module has its own clock, kept by our server rather than your browser, and a closed module stays closed.

Your correct answers in each section are converted to a 200–800 section score using a table for the second module you were routed to, and the two sections add up to your 400–1600 total. We calibrated those tables ourselves — the College Board’s equating data is not public — so every score is reported as a range: ±60 on the total after one practice test, ±45 after two and ±40 after three or more, and ±30, ±25 and ±20 on each section. How scores work explains every step.

A new question starts with the difficulty its author assigned, on a scale of 1 to 5. From the first submitted practice tests onwards, every answer nudges that estimate — a correct answer makes the question look a little easier, a wrong one a little harder — with the nudges getting smaller as answers accumulate. Unanswered questions, answers given too quickly to have been read, and practice tests with integrity problems do not move it.

A question graduates to a proper statistical model only when it has the evidence for one: 150 usable answers for a one-parameter model, 400 for a model that also measures how sharply the question separates stronger from weaker students, and 800 for one that also estimates how often it is answered correctly by guessing — and only if that guessing estimate is precise. A question never moves back down a rung. If a model fits a question badly, the question is flagged for a person to look at; software never retires or rewrites a question on its own.

These models are refitted once a week, from answers read without names, email addresses or account numbers. Today their job is quality control and calibration — finding questions whose statistics suggest a problem. They do not yet change how your score is calculated; if that changes, how scores work will say so first.

For each skill we estimate the probability that you have learned it, using a well-established method called Bayesian knowledge tracing. It starts from values published in learning-science research and updates after every answer: a correct answer raises the estimate, a wrong one lowers it, and both allow for the fact that people slip on things they know and guess right on things they do not.

You see a range, like 58% (34–81%), not a single figure. The range shows how much the estimate would move if the method’s own settings were somewhat different, so it is wide when you have given it little evidence and narrows as you answer more. Nothing is shown for a skill until five answers count towards it.

Once a skill has answer histories from 150 students, its settings are refitted from real data — and kept only if the fit makes sense, for instance that a student who has learned a skill is more likely to answer correctly than one who has not. When that happens the range halves.

• Practice-test answers count only once the test is submitted — never while you are still sitting it
• Lesson and drill answers count, but only your first attempt at a question on a given day
• Practice-test answers given too quickly to have been read do not count
• A practice test later voided for integrity reasons stops counting

For each question there is a time below which an answer is treated as a rushed guess: the larger of three seconds and the smaller of 20% of that question’s typical answer time or 25% of the time the module allows for it. Until a question has 30 answers to take a typical time from, the threshold is 20% of its time allowance. Extra-time accommodations do not change the threshold.

A rushed guess is never marked wrong and never changes a score — a right answer is a right answer. It is simply not used as evidence about what you know or how hard the question is, so a lucky or unlucky click cannot distort either. Your results page shows how many answers were fast and which rule was used.

Every recall question you have answered in a lesson or a drill is scheduled for review with FSRS, an openly published spaced-repetition method. How it went sets the next gap: a wrong answer brings the question back tomorrow, and a right answer pushes it further out — further still if you said you were certain, less far if you said you were not sure. The schedule aims for you to still remember a question nine times out of ten when it comes back, and days are counted on your own calendar.

Questions that are due come first in your drills, most overdue first, and today’s mission says how many are waiting. Everyone’s reviews are timed by the method’s published defaults until they have completed 1,000 reviews; after that, a weekly job can fit the timing to your own history.

Skills within a domain are ordered, and each builds on the one before. A skill you have not practised stays locked until the skill before it reaches 60% with at least five answers behind it — practise a locked skill anyway and it unlocks. A skill counts as mastered at 90% with at least eight answers, and moves to the back of the queue.

Among the rest, drills favour the skills where your recent results are moving away from your longer-term average — that is where practice is paying off, or where something is slipping — with an extra nudge towards skills you have hardly practised, so nothing is ignored just because it has not been tried, and a smaller nudge towards lower mastery. Due reviews come before all of it, and a question you got right in the last day goes to the back.

Each drill question says why it was picked: Due for review, New for you, You are improving here or Worth another look.

Your plan is built from your survey answers, your test date and the estimates above. After you finish a set of drills, complete a lesson or submit a practice test, it checks whether any of that has changed, and if it has, it rearranges itself straight away. There is no daily limit on how often that happens. Lessons are ordered by where your mastery is lowest; drills follow the order described above.

Rearranging never undoes your work: a finished task stays finished and a lesson you have started keeps its place. When the plan does move, your daily mission tells you what changed — for example that a skill moved earlier or a new one unlocked. And it only ever schedules things you can actually open.

Everything above is calculated by our own software, in our own database. An AI model writes the debrief on a practice test’s results page — from a summary of numbers our software has already calculated, with no identifier attached — and the one-sentence explanations you can ask for in lessons, drills and review. In the debrief it may not state a number of its own: the figures are filled in from our calculations, and a draft that contains its own number is discarded.

How we use AI names the model and providers, shows exactly what is sent, and says what is kept and for how long.

The graphing calculator in the Math modules is our own, built to work like the calculator in the real test. We record how many expressions and graphs you made — the database strips anything else from those records on the way in — and never what you typed. Your highlights, notes and scratchpad are saved so a dropped connection cannot destroy them, and are shown to nobody but you.

During a practice test we also record integrity signals, such as the tab losing focus or the clock on your device drifting from ours. They can only affect whether a practice test appears on the public leaderboard — never its score — and a person reviews any appeal. The privacy policy lists every signal.