What the CCAR-F exam measures

What you will be able to do

• Name the five CCAR-F domains and their published weights
• Describe the exam format, length, delivery and fee
• Explain what a scaled score of 720 on a 100-1,000 scale means
• Turn domain weights into a defensible allocation of study time
• State plainly what this guide is and where the authoritative source lives

The Claude Certified Architect, Foundations credential is not a vocabulary test. Its subject matter is four things that ship together in production, Claude Code, the Claude Agent SDK, the Claude Application Programming Interface (API) and Model Context Protocol, and most questions ask a version of the same underlying thing: here is a system that already runs, here is how it is misbehaving, what would you change and why that rather than the obvious alternative?

That framing should change how you prepare. Someone who can define a hook, a subagent and a tool description can still fail, because the exam rarely asks for the definition. It asks which of four plausible repairs fits the symptom, and three of the four are real techniques applied to the wrong problem.

The credential at a glance#

One naming point worth fixing early. The exam code is CCAR-F. Third-party study sites, forum posts and aggregator pages frequently write it as CCA-F, which is close enough to be confusing when you are searching for material. They are referring to the same credential.

Two of these rows matter more than they look. Sixty items in 120 minutes is two minutes each, and scenario-based items carry a preamble you have to read, so the real budget is closer to ninety seconds of thinking per item once reading is paid for. And every multiple-response item tells you how many responses to select, which removes the usual guesswork about whether an answer set is complete but also removes the excuse for a half-formed answer.

Five domains, five weights#

The weights describe how much of the blueprint each domain occupies. If they map evenly onto 60 items, domain 1 is around sixteen items and domain 5 around nine, but treat that as arithmetic rather than a promise: the published blueprint states weights, not item counts.

The useful reading is comparative. Domain 1 is worth roughly half again as much as domain 5, and agentic architecture is also the domain where a wrong mental model does the most damage, because loops, coordinators, subagents, handoffs and session state all interact. Domains 3 and 4 are equal at 20% each and are the two most self-contained: Claude Code configuration is a finite surface you can learn thoroughly, and structured output has crisp right answers. Those are the cheapest points on the paper.

What a scaled score of 720 means#

This is the single most misread number on the page. 720 out of 1,000 is not 72% correct. The scale is not a percentage in disguise, and the raw proportion of items you need is not published, so nobody outside the certification programme can tell you what it is. Anyone who states a specific raw percentage is guessing.

The reason for scaling is equating. Because the exam draws four scenarios from a bank of six, two candidates can sit meaningfully different papers on the same day. Scaling adjusts for that difference so that 720 represents the same standard of competence on every form. The practical consequence is that you cannot calibrate against a raw percentage from a practice test, including ours, and conclude that you are safely over the line.

What the score report gives you#

Results come back as pass or fail with a scaled score, plus percent-correct broken down by domain. That breakdown is the useful artefact, especially after a fail. It tells you whether you missed broadly or lost one domain badly, and those need different responses.

Read the per-domain figures with the likely item counts in mind. If domain 5 carries around nine scored items, one wrong answer moves it by roughly eleven percentage points, so a 67% there may mean three wrong answers rather than a structural gap. Treat that as indicative rather than exact: the blueprint publishes weights, not item counts. Domain 1, with more items behind it, gives a steadier signal.

How the standard was set#

CCAR-F is criterion-referenced. The standard behind it came from a formal standard-setting study: trained subject matter experts worked through the exam content and judged what a minimally qualified candidate should be able to do, and 720 is where that judgement landed on the reporting scale. Scaled scoring then equates across forms, so the same number describes the same competence whichever paper you happen to draw.

The practical reading is that the line is a description of demonstrated skill against the blueprint, not a rank. You cannot be pushed below it by other candidates, and you cannot argue your way over it by being close to the median. Prepare against the objectives, because that is literally what the experts were judging.

The per-domain percentages on your score report are informational. Pass or fail rests on the total scaled score alone. This is worth stating plainly, because the report invites the opposite reading: five percentages sitting next to a verdict look like five hurdles you had to clear individually. They are not. A domain you did badly in can be outweighed by the rest of the paper.

If it does not go your way#

A failed attempt is not the end of the route, but the waiting period lengthens each time.

You may sit up to four attempts per rolling twelve-month period, per exam, and the fee applies to every attempt. Two details in that sentence matter. The period rolls rather than resetting on a calendar boundary, so a fourth attempt in November is governed by what you sat since the previous November. And the limits are per exam, so a failed CCAR-F does not stop you registering for a different exam in the programme.

If you believe something went wrong with the administration or with your result, an appeal goes through Pearson VUE support within 14 days of notification, or within 14 days of the exam date where the concern is about the result itself. Two things sit outside appeal: the outcome of the standard-setting study, so "720 is set too high" is not an argument anyone will hear, and the content of individual items.

Booking, identification and conduct#

Registration and scheduling run through the Anthropic Partner Academy and Pearson VUE, and the fee shown at checkout already reflects any partner-tier discount you are entitled to. You can cancel or reschedule up to 24 hours before the appointment. Inside that window, and for a no-show or a late arrival, the fee is forfeit and you register again from the beginning.

Bring valid, unexpired, government-issued photo identification whose name matches your registration exactly. If you need accommodations, request them through Pearson VUE and have them approved before you schedule, not afterwards.

The exam is proctored. Online, that means staying in view of the proctor and webcam, keeping the workspace clear and communicating with nobody. In every delivery mode it means no capturing, copying, photographing or reproducing exam content in any form. Misconduct is not met with a warning: it can invalidate the result, revoke a credential and bar future attempts.

Keeping the credential#

Renewing on time is a free, non-proctored assessment on the Partner Academy covering what has changed since you certified. Letting the credential lapse means sitting the full exam again at the full fee. The guide does not say when the renewal window opens, or whether any grace follows expiry, so do not plan around one: it states only that renewal is free when it is on time and that a lapse costs a full sitting. If exam content has moved significantly, Anthropic may require the full exam rather than the renewal assessment, so treat the free path as the usual case rather than a guarantee. The cheap move is a calendar reminder a month before expiry.

What "foundations" means here#

Foundations does not mean introductory in the sense of gentle. It means the exam stays at the level of architecture and configuration decisions rather than deep implementation: you are expected to choose between a coordinator and a single loop, to say when plan mode beats direct execution, to recognise a tool description that will cause the wrong tool to fire. You are not expected to recite function signatures.

The credential is valid for twelve months from award, which is a deliberate statement about how fast this material moves. Treat anything you learn here as perishable, and check the current documentation for the products before you rely on a specific flag, field or file path.

The next lesson covers the six production scenarios the questions are framed inside, because knowing the settings in advance is most of what makes an unfamiliar item readable under time pressure.