How our exams are made

We first derive the exam’s domain blueprint — the 6–16 topic areas the real exam actually covers, with their approximate weightings (for AWS Solutions Architect that means areas like compute, storage, networking, security). Every practice set for a subject is generated against this same blueprint, so coverage stays consistent and balanced instead of drifting set to set.

Generation runs under strict originality instructions: never reproduce or paraphrase real exam questions, question banks, or “braindumps.” Each question must be a freshly invented scenario with its own names, numbers, and context. A per-set uniqueness seed pushes every generation run toward different phrasing, so two sets never converge on the same wording.

This matters for two reasons: copying real exam content is a violation of certification NDAs and does not actually teach — and we want neither.

After generation, every question is compared against every other question in the set and against all earlier sets for the same subject, using a text-similarity check. Anything too close to an existing question is dropped before you ever see it.

Every question ships with a full explanation — why the right answer is right and why each distractor is wrong — plus links to official documentation (vendor docs, exam guides, standards bodies) where the concept is covered.

Then we actually fetch every link before the question reaches you, and drop any that don’t resolve. This matters more than it sounds: when we first measured it, a large share of generated links were plausible-looking but dead — correct-looking paths on real sites that simply never existed. A model is good at the shape of a URL and bad at the path. Checking is the only thing that catches that, so we check.

Each question is then checked against published multiple-choice item-writing guidelines — the taxonomy from Haladyna, Downing & Rodriguez (2002) and the checkable rubric derived from it by Tarrant and colleagues (2006). These catch the flaws that let someone score well without knowing the material: the correct answer written longer and more carefully than the alternatives, a distinctive word shared only by the question and the key, “all of the above,” duplicated options.

Questions with defects that make them unanswerable are removed. Style warnings are recorded and shown to you rather than hidden — you can see the counts for any set on its exam page.

For each generated set we permanently record which model produced it, which version of our generation instructions was used, when it was created, and the provider’s request ID. That log exists so we can always demonstrate a set was independently created — and trace exactly how any question came to be.

Where an exam has an official cut score (AWS ≈ 72%, CompTIA Security+ ≈ 75%), that’s the pass mark we show. We don’t inflate difficulty to flatter you or deflate it to sell you.