Three AI models check each other's exam questions, with no human reviewer

A developer building practice exams for Dutch certifications, including workplace safety, boating and taxi licences, says the questions are drafted by a language model and never read in full by a person. Instead, each question goes through a pipeline in which a second and then a third model check it blind, with one redraft allowed for failed items. A rule enforced in code requires the verifier to be a different model from the first checker, and a test confirms that cited source quotes appear verbatim. Across roughly 10,000 check results, the third model failed 14.3% of questions the second had passed, most often for being off-level, having a flawed explanation, or being ambiguous rather than having a wrong answer key. The author also reports a script error that broke 62 questions by reassigning answer letters after explanations were written, which the checks caught.