Sources and checks

What every question's answer rests on, when each exam was last checked against Anthropic's documentation, and the safeguards designed into the questions.

What each answer rests on

Every question records what its answer rests on: Anthropic's documentation, the exam guide, or established practice such as GDPR. After you answer, Read more links under the explanation take you to those pages. Where the exam guide and today's documentation differ, the explanation says so, and the study guide's Exam guide vs today page lists the differences recorded for each exam.

• Foundations (CCAR-F): last full check 14 October 2026, against Anthropic's current documentation and exam guide v1.0 (July 2026). Last checked against Anthropic's releases 24 October 2026. Questions last added 5 October 2026 (21 new), each checked against its cited sources before release.
• Professional (CCAR-P): last full check 14 October 2026, against Anthropic's current documentation and exam guide v1.0 (July 2026). Last checked against Anthropic's releases 24 October 2026. Questions last added 5 October 2026 (21 new), each checked against its cited sources before release.

The exam guides only change when Anthropic publishes a new version, but its documentation changes all the time, and many answers rest on it. So both exams are checked in three ways:

• Against Anthropic's releases, every two to three weeks: the release notes since the last check are read, and the questions and study guide pages that a change touches are checked again.
• In full, every quarter: every question goes through an independent review against the sources it cites and today's documentation.
• Against each new exam guide, when Anthropic publishes one: the format, the domain weights and every objective are compared with it, followed by a full check.

New questions are checked against their cited sources before they go in, and the next full check covers them again.

No questions from the real exams

ClaudeArena's questions are written from Anthropic's published exam guides and documentation, never from the real exams. Candidates agree to keep the exam confidential, and Anthropic's exam policy treats sharing real exam questions, or seeking them out, as misconduct that can cost a candidate their certification. So if you've sat the exam, please don't send us its questions, in a report or anywhere else. A question reported as coming from a real exam is taken out of the bank, not reviewed.

Question-design safeguards

Fact-checking is one part of a fair practice bank. These are designed into the questions and checked by the software before every release:

• Blueprint coverage: full exams follow the published domain weights and every question is tagged to an exam objective. Foundations scenarios rotate by type and version rather than repeating one fixed exam.
• No shortcuts from the answer choices: options are shuffled for every sitting. The build also asks whether someone who never reads the question could beat guessing by looking only at the shape of the options: their length in characters or words, or how many commas, hyphens, dashes, brackets, colons, semicolons or slashes they carry. It tries the simple tricks, such as always picking the longest, or picking differently when the longest or the shortest stands out, on every exam and mode, for choose-one and choose-two questions, and fails if any of them does better than chance alone would explain. A mark that only one option carries, such as brackets or quotation marks, mustn't point to the right answer or let you safely rule that option out, and neither may the one option without a mark all the others share. Within a question, options end the same way. The aim is that no such shortcut is worth using, not an artificial exact 25% for every position. Word choices, such as an option that joins two actions with "and", aren't checked yet; that work is planned.
• Plausible distractors: wrong answers are written as specific, credible mistakes or trade-offs, not filler added to change their length. An automated check rejects repeated phrases that suggest padding, and every content change is still read against the question's explanation.
• Server authority: the server draws questions, shuffles choices and grades answers. The page never receives the full confidential bank, and it receives an answer and explanation only once that feedback has been earned.

More about ClaudeArena: About ClaudeArena · How it works · Your data.