A training platform is worth what its questions are worth. Rather than assert it, this page publishes what we measure on our bank: the defect we found, how we fixed it, the permanent check that prevents its return, and what is not up to standard yet.
Every figure on this page comes from a dated internal measurement; the last verification date is at the bottom of the page.
Displaying a figure creates an obligation: keeping it accurate. That is exactly why we do it. A candidate preparing a selective competition deserves to know what they are training on, and an indicator published with its date is worth more than any adjective.
To our knowledge, no competing platform publishes a measured quality indicator for its question bank. We prefer to show ours, defects included.
A classic MCQ defect, documented by item-writing research: when the correct answer is more developed than the wrong ones, always ticking the longest option earns points without any knowledge. We measured what that strategy scored on our English questions, module by module, then rewrote the distractors of 690 source questions, with a complete line-by-line review.
| Module | Before | After | Chance |
|---|---|---|---|
| Digital skills | 86.5% | 26.9% | 25% |
| Verbal reasoning | 68.5% | 29.3% | 25% |
| EU knowledge | 64.7% | 30.2% | 25% |
Numerical reasoning, whose options are numbers, did not carry this bias. Residual measured on 2 August 2026 across the 8 primary languages: from 22.0% to 31.9% depending on the language/module pair, over 44 measured pairs.
A correction without a check degrades with the next batch of content. Length bias is therefore measured every week by an automated continuous-integration check, on every language/module pair of the 8 primary languages. If the longest-answer strategy exceeds 35% (chance is 25%), the check turns red and blocks: the threshold is calibrated to tell a real defect from statistical noise.
The same continuous-check setup locks other content invariants: parity of the eight interface locales, consistency of test parameters with the official database, uniqueness of FAQ questions across pages.
Publishing measurements also means publishing what they reveal. Known state as of 4 August 2026:
One-off corrections are not enough without writing rules. These apply to every new question:
Get upcoming competition notices and calendar updates.
Sign-up opens soon.