How Ninetynine's CAT data is built

The questions and answer keys are CAT's own and are never changed. Worked solutions are written blind, then accepted only if they reach CAT's answer. Topics, concepts and similar questions are Ninetynine's own work, described below so you can judge them.

Where do the questions come from?

From the 24 CAT papers held between 2017 and 2025, 1,830 questions in all, as transcribed from the papers released after each exam. Figures are not reproduced on this site; questions that depend on a figure say so.

Are the answer keys changed?

No. The answer shown is always CAT's published key. While writing solutions we found a handful of questions that are ambiguous, or whose key does not match their own wording. We keep CAT's key anyway, because that is what was scored.

How are the worked solutions written and checked?

Each solution is written without looking at the key, step by step, with a shortcut where one exists. It is then checked automatically against CAT's answer, and is published only if it reaches the same answer. Solutions are drafted with the help of an AI model (Claude, by Anthropic) and checked by that rule. 893 of 1,830 questions have a published solution so far; the rest show the answer only, and are kept out of search results until solved.

How are topics and concepts assigned?

Every question has exactly one topic from a fixed list of 34 (16 for QA, 9 each for DILR and VARC); RC questions are tagged by the subject of their passage. Concept clusters go further: they group questions from different papers that need the same method, such as "remainder of a large power: find the cycle". Each member carries a similarity score, and near-repeats (same setup, new numbers) are flagged. There are 186 clusters; the 143 with three or more questions have their own page.

How are similar questions made?

A similar question tests the same concept with the same method, but with new values and new wording. It is solved again from scratch without its key, every option is checked so exactly one is right, and counting or number-theory answers are brute-forced by a short program. For DILR the unit is a whole set: a new caselet whose facts are encoded and every possible arrangement enumerated, to prove each question has exactly one answer. Similar questions are for signed-in practice and are not published here.

How are the numbers in the guides computed?

By a script that reads the papers, tags and clusters at every build. No number on a guide or topic page is typed in by hand, so they stay right when the data changes. Counts are questions, not marks.

Found a mistake?

Tell us. We check every report against the original paper.