How marking works

1. A mark scheme comes first. If you paste the official one, that is used. A practice paper built here already has one, written with the question before anyone answered it. If there is none, one is written from the question before your answer is read, so the scheme can’t be bent to fit what you wrote.
2. Your answer is read against it, point by point. Short answers earn marks for separate creditable points. Long answers (usually 6 marks and up) are judged against levels of response, the way Cambridge does it.
3. The mark is added up, not guessed. Your total is the sum of the points that were ticked, so the number always matches the ticks you can see. You can open the mark scheme under any question.
4. Doubt is flagged. If the handwriting was hard to read, or the level was a close call, the mark is labelled “double-check” instead of being presented as certain.
5. Grades are arithmetic. A grade is never asked of the AI. It is your percentage read against Cambridge’s published grade thresholds for that syllabus (Core or Extended where the subject has tiers) and the page says which.

We wrote 13 answers to Cambridge IGCSE Economics questions at known quality levels (full-mark answers, half-right ones, confident nonsense, and deliberate traps) and marked each one several times with the same settings the site uses.

• With a mark scheme: the mark matched the reference exactly 85% of the time and was within the allowed tolerance 100% of the time. Before answers were marked against a scheme, those figures were 44% and 79%.
• Same answer, same mark: at the setting we run in production, every one of the 13 answers got an identical mark on every run. Consistency is the thing a chatbot is worst at, and the reason we measure it.
• A trap it now catches: an answer that only named two reasons, without explaining either, used to get 4/4. It now gets 2/4, every time.
• Without a mark scheme it is weaker: between 82% and 95% within tolerance, and less consistent from run to run, because the scheme it writes for itself varies. Pasting the official mark scheme is the single biggest thing you can do for accuracy.

• The reference marks were set by us, marking carefully to Cambridge’s conventions, not by a Cambridge examiner. This is not a certified accuracy figure.
• Thirteen answers in one subject is a small test. It tells us the method works and catches regressions; it does not prove every subject marks equally well.
• We have not yet tested against real scripts marked by real examiners. When we do, the results go here, good or bad.

• Long, levels-marked answers: 8-mark “Discuss” questions and History essays. Two human examiners can differ by a mark or two here; so can we.
• Messy handwriting and diagrams. It reads what it can and flags what it couldn’t. Typing the answer removes the doubt.
• Questions with no mark scheme, as above.
• Practice questions are AI-written in the style of past papers. For the real thing, bring a real past-paper question and its mark scheme.

A mark from GradeGenie is a guide to improve with, not a prediction of your result. If a mark looks wrong, tell us from the feedback button on the result; we read every one.

The quickest test is your own work: take an answer your teacher has already marked, upload it with the mark scheme, and compare. Create a free account: about 15 marked answers a month cost nothing.