How every question is checked
A wrong answer key teaches you something false. So no question reaches you until it has passed every check below. When a question fails one, it's thrown out. When too many fail, we tell you instead of padding the set.
What the checks can't promise
Is the material readable?
A badly scanned page reads as broken fragments. Rather than invent questions from it, we ask for a clearer copy.
Write more questions than needed
The writer drafts about half again as many as you'll get, because some won't survive the checks.
Every question quotes your notes
Each one must copy, word for word, the line its answer comes from. The program checks that the line is really there. A question whose quote isn't in your notes is thrown out.
Rule checks on every question
Four different options, exactly one right. No "all of the above". No trick "which is NOT" wording. No answer given away by grammar, by the word "always", or by being the longest option. Any number in the answer must appear in its quote.
A second reader answers without the key
A separate AI call gets the question, the options and the passage, but not the answer. If it picks a different option, says two could be right, or can't find the answer in the passage, the question is thrown out. On the Exam plan this happens three times, with the options shuffled each time.
Disputed questions get one more look
On the Exam plan, a disputed question goes to a stronger model, which may repair it. The repair has to pass every check again.
Keep the good ones, or say so
The right answer's position is shuffled so you can't learn a pattern. If too few questions survived, you get an honest message and the set doesn't count against your month.
Keep a record, not your notes
Each run records how many questions passed and why others failed. It never records your text. Those counts are how we find and fix a weak spot.
The checks prove a question matches your material and that an independent reader agrees with its answer. They can't tell whether your material is right. If your notes say the wrong date, a faithful question will ask for the wrong date. That's why every question is labelled with where it came from, and why you can always see the line.
Questions made from general knowledge, like the topic quizzes in the break games, have no notes to quote. They're labelled as general knowledge and still go through the rule checks and the second reader.
No method makes mistakes impossible. We measure how often they slip through and publish changes in the changelog.