CCAR-P - Solution Design & Architecture (17% of the exam) - Section 1.2

Design end-to-end architectures (input, processing, output, feedback loops).

The full path of a Claude system: how input arrives and is validated, where processing and model calls sit, how output is checked and delivered, and how feedback from users or evaluation flows back into the system. Candidates should spot the stage an architecture is missing, most often the feedback loop.

input validationprocessing and model invocationoutput validation and deliveryfeedback loops

Practice question for this objective

Free sampleSolution Design & Architecturemedium

A retail bank uses Claude to extract declared income from payslips that applicants upload during a mortgage application. After the bank launched a mobile upload route alongside its existing web form, the share of extractions that underwriters correct rose from 2 to 9 percent, and 85 percent of the corrected cases arrived through the mobile route. The prompt, model and output schema were not changed in that release, and every extraction still passes schema validation. Where should the architect look first?

AThe model tier: whether a more capable model reads low quality phone photographs more accurately than the current model does.
BThe input stage: what the mobile route actually submits, such as page count and image legibility, compared with web uploads. Correct
CThe output schema: whether its field types are loose enough to let a plausible but incorrect income value pass validation.
DThe prompt: whether its extraction instructions cover the newer payslip layouts that some employers have started to issue.

When errors rise after a change to how input arrives and concentrate in the new channel, investigate the input stage before the model or prompt. The 'what changed' evidence isolates the input stage: prompt, model and schema are unchanged, and the errors are concentrated in the newly added upload route. Phone uploads commonly differ from web uploads in ways that affect extraction, such as cropped or missing pages and low legibility. Validating input completeness and quality at the point of arrival stops the model from extracting confidently from a document that does not contain the income figure, and schema validation on the output cannot detect that kind of error.

Why A is wrong: This is tempting because a stronger model may cope better with poor images. It is wrong as a first step because the model did not change and the failure followed a change to the input path; upgrading the model compensates for bad input instead of finding out what the new route sends, and cannot recover a page that never arrived.

Why B is correct: This is correct because the only thing that changed is how documents arrive, and the errors concentrate in the new route. Comparing what the mobile route delivers, for example missing pages or blurred photographs, with web uploads tests the most likely cause directly, and points to input validation that rejects or flags incomplete documents before the model call.

Why C is wrong: This is tempting because every extraction passes schema validation and yet some are wrong. It is wrong as the first place to look because a schema only checks shape and type, it did not change, and it cannot explain why the errors cluster in mobile uploads; tightening it would not address the cause.

Why D is wrong: This is tempting because unfamiliar document layouts are a common cause of extraction errors. It is wrong because new employer layouts would reach the bank through both upload routes equally, whereas 85 percent of the corrected cases came through the mobile route, which points at the channel rather than the document design.

See more CCAR-P practice questions, answers explained.

Exam traps in Solution Design & Architecture

Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-P bank for this domain.

Add a system prompt instruction telling the model to state clearly in each summary which mandatory documents appear to be absent
Why it is wrong: This is tempting because it is a one-line change that makes missing documents visible to officers. It is wrong because the summary is still drafted from the incomplete application, which the requirement forbids, every such application still consumes the fixed budget, and a prompt instruction is a soft control the model can miss.
The model is reaching its maximum output length on long answers, so raising the output token limit will restore the endings.
Why it is wrong: This is tempting because answers that stop mid-sentence are a familiar sign of a response hitting its output limit. It is wrong because the application logs hold the complete text as the model returned it, so the model finished each answer and the loss happens later in the path.
Clinicians are rewriting drafts to suit personal style, so the edit rate tracks preference rather than the accuracy of the content.
Why it is wrong: This is tempting because edit rates on generated drafts often include stylistic rewording that says nothing about accuracy. It is wrong here because the sample of 60 heavily edited drafts shows the corrections are substantive medication changes, which is a content failure rather than a matter of taste.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.