Why a question from this site is safe to study
With two weeks left you cannot afford to memorize something a model made up, so here is exactly what has to happen before a question reaches you: the quote has to exist, the page has to match, and a second model has to reach the same answer without seeing the key.
Every accepted item prints its evidence the way this one does: "Preload is the end-diastolic volume that stretches the ventricular myocardium before contraction.", page 1. That sentence was found in the stored text of page 1 before the question was kept, and a second model from a different family then picked Preload from the quote alone, without the key.
Free to try. No card and no account needed.
54 percent of the questions drafted since counting began on failed a check and never reached a student. Counted from this site's own database, as of , refreshed twice an hour. A question fails when its quote is not on the page it cites, when a second model picks a different answer, when it breaks a writing rule or when it repeats one already written. Each check, with its own rate.
Try the sample exam, ten real questions with no upload and no account.
What no quiz generator does
Every question carries the sentence it was written from, copied out of your file, with its page number.
Before you see a question, the quoted sentence is matched against the stored text of the page it cites. If it is not there, the question is dropped.
A second model from a different family answers the question from the quote alone, without the key. Disagreement drops the item.
Options like all of the above, vague which of the following is true stems, and questions about "the text" instead of the subject are rejected by pattern before they reach you.
Each new item is compared against everything already generated from the document, so a long run gives you different questions instead of twelve facts restated.
Two of these five are the checks that reject on evidence, the quote check and the blind second answer. The other three are rules about what a question must carry before those checks run.
Works in any language, and quality varies by language: the questions come out in the language of your document, or in the one you pick from a list of every living language. What works in your language, and what does not.
What happens to one question
A textbook chapter, a PowerPoint deck, a Word handout, a photographed set of notes, a plain text file, text you paste into the box, or a public web page by its address. Text PDFs can be read in your browser and imported in seconds without uploading the file. Large files upload in parts, so a 900 page book is fine.
The server stores the text of every page separately, with its number. A PowerPoint file is stored one slide per slide. Word files, text files, pasted text and web pages have no pages of their own, so they are cut into numbered sections of about 2,500 characters. PDF pages with no text layer go through optical character recognition, so scans and photos end up with the same page-accurate text as digital PDFs.
A model receives a slice of the document and writes questions of the types you asked for: single best answer, clinical vignette, multi-select, true or false, short answer. Each question must come back with one or two verbatim quotes from the slice and the page number they sit on.
The server looks for each quote in the stored text of the page the model cited, after normalizing whitespace and typography. If the quote is not there, the question is discarded. This is what stops invented facts and invented page references.
A second model from a different family is shown only the question and its quotes, never the answer key. It answers on its own. If it disagrees with the key, the question is discarded. Ambiguous questions fail this step as often as wrong ones do, which is the point.
Answers are graded automatically, short answers by a grading model against the source. You get an A to F grade, mastery by chapter, a review queue of what you missed, and a tutor that can only argue from your document.
What a hallucinated question looks like, and which check catches it
A hallucinated question is usually one of three things: a fact the document never stated, a quote the model wrote itself, or a real quote attached to the wrong page. The quote check handles the last two, because a line that is not in the stored text of the cited page cannot pass. The blind check handles the first, because an answer key the source does not support is one the second model has no reason to pick. Neither check makes the output perfect, which is why every question keeps its quote and page number: you can settle a disagreement by reading the line.
The checks, and what each one throws away
A tool that cannot tell you what it rejected has told you nothing about what it kept. Rejection is normal and it is the reason the bank is worth studying from. A typical run discards questions for a misquoted line, a quote that belongs to a different page, an answer key the second model will not agree with, or a near duplicate of a question already accepted from the same passage. You only see what survived, with its sources printed underneath.
How much each model family loses to those checks is published, over the last 30 days and with the number it is counted from: rejection rates by model family. The structural rules a question has to pass before any of that runs are in your browser at the MCQ flaw checker, where you can paste a question of your own.
Measured on real files
| Lane | File | Work | Wall clock |
|---|---|---|---|
| Free pool | An OpenStax anatomy chapter of 44 to 60 pages, text PDF | 20 questions generated and checked | about 15 to 18 seconds |
| Reading only | A 900 page textbook, text PDF | every page read and stored | under 40 seconds |
Measured on the live site, the chapters on September 23, 2026 and the textbook on September 12, 2026, from the moment the job started to the moment the last question was accepted. Your own times move with the length of the job and how busy the free lane is that hour.
The free lane writes on DeepSeek V4 Flash first and hands a call to another open-weight model when that one is busy, so a large job keeps moving. The output is the same shape and passes the same two checks whichever model wrote it. The runs in the table above are the only timings this site claims, and each one names the file and the date it was measured on.
Cost
What the pass changes is the models, the unlimited tutor and the downloads. What it does not change is the quote check and the blind second answer, which every question passes on either lane.
- Free: 300 verified questions a month, on the same two checks as every paid plan.
- Exam Pass, $19.99 once: the priority models for the 90 days of the pass, which are stronger on long clinical vignettes.
- A question pack of 500 questions, $5.99 once: the same models and the downloads with no subscription, and a question the checks throw out is not charged.
See pricing for the rest.
Limits worth knowing
- A question can only be as good as the page it came from. A vague lecture slide gives vague questions.
- Diagrams and image-only figures are not read as images. Their captions and surrounding text are.
- The tutor is restricted to your document. If the answer is not in the text you uploaded, it says so instead of guessing.
- Nothing here is medical advice. It is exam practice built from material you chose.
Who builds this
One person writes the code, answers the email and decides what the checks reject. That is why the pipeline above is described in this much detail rather than summarized as a promise: there is nobody here to appeal to if a question turns out to be wrong, so the page each question quotes has to be enough to settle it yourself. What has changed and when is on the changes page, and the inbox is on the contact page.
No account needed to try it. Registering keeps your library, grades and review queue across devices.
Published , updated .