Methodology
How we write our questions — and how we throw them away
Exam prep is easy to sell and hard to verify. A bank of ten thousand questions looks better than a bank of three thousand, and nobody can check whether the answers are right until they've already paid. So here is exactly how ours are built, and what we do when they don't hold up.
Start from the published outline, not a textbook
Every state's testing vendor publishes a content outline: the topic areas and how many scored questions each carries. We build to that structure directly, so the proportions you practise match the proportions you'll sit. When a state weights contracts at 19 of 100 scored items, that's what the bank reflects.
Write questions to the blueprint
Questions are drafted with AI against that outline, under an explicit rule set for the state's own law — the terminology it uses, the forms it requires, the doctrines it doesn't recognise. Maths questions come from parameterised templates instead, so the arithmetic is computed exactly and the distractors are real mistakes rather than random numbers.
Review every question with two independent models
This is the step most prep material skips. Each question is examined separately by two different AI reviewers, prompted as state-law experts and told to hunt for wrong answer keys, factual errors, ambiguity, and explanations that contradict the marked answer. They don't see each other's verdicts.
Delete anything both models flag
Where both reviewers independently flagged the same item, we deleted it — not softened it, not re-worded it. Across the six states put through it, that removed 1,340 questions from 11,550 reviewed. We would rather ship a smaller bank you can trust than a bigger one you can't.
Recalibrate from how candidates actually perform
Once a question has been answered enough times, its observed difficulty is compared to its assigned difficulty. Items that behave oddly — everyone gets them right, or nobody does — get flagged for another look and can be retired from serving.
What the review actually found
Questions removed because bothindependent reviewers flagged them. A single reviewer's objection wasn't enough to delete — that bar keeps the process honest in both directions.
| State | Reviewed | Deleted | Rate |
|---|---|---|---|
| Texas | 3,140 | 331 | 10.5% |
| Pennsylvania | 1,100 | 102 | 9.3% |
| North Carolina | 1,184 | 158 | 13.3% |
| Nevada | 1,212 | 73 | 6.0% |
| Washington | 1,774 | 361 | 20.3% |
| Florida | 1,348 | 315 | 23.4% |
| California | 1,792 | 0 | 0.0% |
| Total | 11,550 | 1340 | 11.6% |
The failures clustered rather than scattering, which is what made them worth fixing at the source: confusions between escheat and tax foreclosure, homestead exceptions, agency duties owed to a principal versus a customer, and options where two answers were defensible. Each cluster became a rule in the authoring prompt, so the next state's bank started cleaner than the last.
What we hold ourselves to
- Every question is original. We never reproduce actual exam items.
- We cite the official handbook for exam facts, and tell you to confirm with your state commission where rules change.
- We disclose that questions are AI-drafted and AI-reviewed under human-written rules.
- We delete questions we can't stand behind, even when it shrinks the bank.
- We correct errors when you report them, and we publish how to report them.
Common questions
Are these real exam questions?
No. Real exam items are confidential and copyrighted, and anyone selling them is handing you a liability rather than an advantage. Every question here is original, written to the format and content of the published outline.
Are the questions written by humans or AI?
Both, and we'd rather say so plainly than imply otherwise. Questions are drafted with AI under human-authored rule sets built from the official outlines and statutes, then screened by two independent AI reviewers, with human decisions at the gates — what the rules say, what counts as a failure, and what gets deleted. Maths questions are generated from human-written parameterised templates.
Why does the bank get smaller after review?
Because that's the point of reviewing it. A first pass of AI-generated licensing content had a substantive error rate around 21%. Hardening the authoring rules cut that sharply, and the two-model review then removed the residue — 1340 questions in total. Bank size is easy to inflate; accuracy isn't.
What if I find a question that's wrong?
Tell us at info@agentexamcoach.com and include the question. We correct or remove it. No review process catches everything, and we'd rather hear about the misses than pretend there aren't any.
See the questions yourself
Free practice questions with the full explanation on every one — no account, no card.
Try free practiceSee also our editorial policy and why this exists.