21% of the exam — about thirteen of the sixty questions — and the largest domain on it. That weighting is the exam telling you something: this credential is not mostly about getting Claude to produce things. It is about deciding whether what came back can be used.
The organizing idea, and the one the questions keep returning to:
Accurate is not the same as usable.
An output can be true in every checkable particular and still be the wrong thing to send. Several questions in this domain describe exactly that situation, and the tempting answer is always the one that checks the facts again. The facts were never the problem.
So there are four separate questions to ask, and they fail independently.
Is it true?
The first check, and the one people already know to do. Two things make it harder than it sounds.
Confidence is not evidence. Claude states things in the same steady tone whether it is certain or reconstructing. A number delivered plainly and a number delivered plainly are indistinguishable, and one of them may be wrong. You cannot use how it sounds as a signal, because the sound never varies.
A citation is a claim, not a check. When an output names a study, a journal, a year and a page, that is four more things that might be invented rather than four reasons to relax. The exam likes this scenario, and the credited answer is always to go and look — never to treat the specificity as reassurance.
The practical rule: check against the source, not against the model. Asking Claude "are you sure?" is not verification. It has no new information between the first answer and the second, so what you get back is another answer of the same quality, often more confidently phrased. If the vendor published an annual report, the annual report settles it. If a regulation has a commencement date, the statute settles it. The strongest evidence is the primary document, every time.
When you cannot check everything, and often you cannot, check by consequence rather than by order. The claim that will be quoted to a customer, the number going in front of the board, the figure someone will act on — those first. Checking the first third of a brief because it is the first third is the wrong division of limited time.
Is it faithful?
This is the check people skip, and the exam knows it.
Summarizing is where it bites. You gave Claude thirty pages and got back two that read beautifully. The two pages are internally consistent, fluent, and plausible — and none of that tells you whether the summary kept what mattered. A summary can be perfectly accurate about everything it mentions and still be wrong, because of what it left out.
Nothing inside the summary can reveal this. You have to go back to the source. Not all thirty pages necessarily, but enough to answer: did the things that would change a decision survive? If a grant report contained one paragraph about a funding condition, and the board summary is silent on it, the summary is not a summary. It is a different document that happens to be shorter.
Same shape when Claude works from documents you supplied. Fluency comes from the model. Grounding comes from the source. They arrive looking identical.
Is it for the right person?
Here is where accurate outputs fail, and where a lot of this domain's weight sits.
An explainer about a new overdraft policy can be correct in every clause and useless to members who are new to banking. A patient letter about appointment changes can be precise and still land badly on someone anxious. A draft written in the register Claude defaults to — even, measured, slightly formal — is not automatically the register your reader needs.
From the free practice form
Northgate Credit Union needs a one-page explainer on a new overdraft policy for members, many of whom are new to banking. Claude's draft is accurate but leans on terms like 'discretionary coverage threshold'. The associate has one afternoon before it goes to print. What is the best next step?
That pattern — a distractor that offers to re-check facts that were never in doubt — shows up throughout this domain. When a question tells you the output is accurate, believe it. The problem is somewhere else, and the answer that re-examines accuracy is there to catch people who only have one check.
Can the next person use it?
The last check is about form rather than content, and it is easy to miss because the content is fine.
If eight carriers compared across six criteria are going to be sorted and filtered by colleagues in a spreadsheet, then prose describing the comparison has failed even if every figure is right. It needed to arrive as a table. If someone needs a fact in the middle of a meeting, a formatted document is worse than a sentence in the chat, because the format costs time the situation does not have.
Ask what happens to this output next. Somebody sorts it, pastes it, forwards it, reads it aloud, or acts on it in the next thirty seconds. The right form follows from that and not from what looks most finished.
When a human has to look anyway
Some outputs need review even when everything above passes. Two situations come up repeatedly, and they are worth memorising because they are the answer to more than one question.
When the output affects someone's rights, money, health or employment. A screening question set, a benefits letter, a clinical instruction, a decision notice. The cost of being wrong is not symmetrical with the cost of checking.
When the output could be discriminatory, and the model cannot see it. If a draft asks candidates about family commitments, the problem is not a factual error Claude could catch on re-reading. It is a legal and ethical failure that requires someone who knows what may not be asked. Claude will not flag it reliably, because it does not know your jurisdiction, your policy, or what your regulator did last year.
Accuracy checks do not catch either of these. They are a different kind of review by a different kind of reviewer, and the exam expects you to know when to escalate rather than iterate.
What to carry into the exam
Four checks, and they fail independently:
- True — against the source, never against the model, consequence first.
- Faithful — did the summary keep what mattered? Only the source can say.
- Right reader — accurate and incomprehensible is still a failure.
- Usable form — what happens to this next?
And one habit: when a question tells you the output is already accurate, stop looking at accuracy. The credited answer is somewhere in checks two, three or four, and at least one distractor will be offering to verify the facts one more time.