Where the puzzles come from
The bank holds 115 items, and every one of them is made by code in this site's repository. A generator takes a rule from a library and a seed number, builds the question and the wrong options, and stores the rule with a written explanation. The libraries cover matrices (shapes that change by count, size, shading, position or turn along rows and columns), number series, letter series, spatial rotation, ordering puzzles and syllogisms.
Nothing is copied or adapted from Raven's matrices, Mensa tests, the Wechsler scales, ICAR or any other site. Puzzle types of this kind look alike because the ideas behind them are old, not because a page was taken.
One correct answer, checked by a second program
For each kind of puzzle a separate verifier reads only the question and the options, never the generator's key, and works out which options are right. An item enters a test only when exactly one is.
- Matrices: every rule that could explain the grid is tried on each attribute, along rows and along columns, and exactly one option must follow from some rule.
- Rotation: the options are reduced to a canonical form, and exactly one is the original shape turned, never unturned and never mirrored.
- Series: every rule in the library is fitted to the given terms, and all fitting rules must predict the same next term.
- Ordering and syllogisms: every arrangement that fits the statements is listed, and syllogisms are checked under both readings of an empty group.
A blind check by an AI model
On 11 November 2026 an AI model (Codex) solved all 115 items from text descriptions, without the key. It agreed with the key on 89 items, disagreed on 25 and called one ambiguous. We looked at each of those 26 ourselves.
For the 12 rotation items a separate script confirmed that the keyed option is the only turned copy of the shape; the model had misread figures described in words. The other 14 (8 matrices, 3 letter series and 3 orderings) we solved by hand. The key was right in every case, and the model's answer broke a stated rule each time. No item was dropped.
This is a second opinion, not the proof. The verifiers above are what establish that each item has one right answer, and the model is known to be weak on spatial items.
The forms
Each form is a fixed list of items in a fixed order, the same for everyone. A full form has 12 matrices, 6 series (4 number and 2 letter), 6 spatial rotation items and 6 logic items. Items are spread over five design levels, with 5, 7, 8, 5 and 5 items, the areas interleaved and the difficulty rising through the form.
| Form | Items | Time | Use |
|---|---|---|---|
| A | 30 | 25 min | The main test |
| B | 30 | 25 min | A retest: shares no item with A |
| Quick | 15 | 10 min | The short version |
How the result is worked out
A skipped item counts as wrong, and a wrong answer costs nothing more than a skipped one, so answering is never worse than leaving a blank. The result shows these figures and no others:
- Score: the number of correct answers, and the time used.
- Likely range: where the score would probably fall on a parallel form. With n items and a share p correct, it is n·p ± √(n·p·(1−p)), rounded to whole items. It describes the luck of the draw only.
- Guessing average: the score pure guessing gives, the sum of 1 divided by the number of options over all items.
- Level reached: the highest of the five levels where that level and every level below it have at least two thirds of their items right.
- Areas: correct answers out of the total for matrices, series, spatial rotation and logic.
Why there is no IQ number
An IQ is a rank. It says where a score sits among the scores of a tested group of people, scaled so that the average is 100. This site has no such group yet, so any rule that turned a raw score into an IQ would be made up. We print no IQ and no percentile for our own test.
What IQ and percentiles mean is explained on the IQ scale page, and the percentile calculator converts between the two for the standard scales.
What the test cannot tell you
The five difficulty levels come from the generators' rules, the number and kind of rules in an item. They have not been measured on people, so level 4 means harder by design, not that a certain share of people fail it.
The likely range covers chance only. Sleep, noise, a small screen or a rushed half hour move a score too. This is not a clinical or diagnostic test and does not replace an assessment by a licensed psychologist.
Corrections
If a puzzle seems to have two right answers or none, write to support@pyxl.pro with the form and the question number, or the result link. The item is checked again and replaced if you are right.