CCAR-P sample question: where an end-to-end pipeline loses information
An original sample question by CERTain·exam, based only on Anthropic's public documentation and the public exam guide. CERTain·exam is not affiliated with Anthropic.
Domain 1 · Solution Design & Architecture (17% of CCAR-P in the exam guide effective until December 6, 2026; 18% from December 7, 2026). Objective covered: "Design end-to-end architectures (input → processing → output → feedback loops)."
This question tests whether you look at the whole pipeline before fixing the step that looks guilty. The output is wrong in one specific way; find the stage where that information disappears. Work through it before you reveal the answer.
The question
Ferrowmere Grid, an electricity distributor, uses Claude Sonnet 5.5 to turn substation inspection reports into maintenance tickets. Each report is a PDF of 6 to 10 pages: written notes, a thermal chart whose hot spots show as color bands, and photos where inspectors have circled defects by hand, without labels. A preprocessing script extracts the plain text from each PDF and sends it to Claude with the ticket instructions. Engineers compared 150 tickets with the original reports. Every defect described in the notes appeared in a ticket, but 31 of the 38 defects shown only in the charts or photos were missing.
Which change should the architect make? (Select one.)
- A. Add a list of the defect types inspectors report to the ticket instructions, with a short worked example of each.
- B. Add OCR to the preprocessing script so that any text found inside images is appended to the extracted plain text.
- C. Send each report to Claude as a PDF document, so every page arrives as an image along with its text.
- D. Split the extracted text of each report into sections and process every section in its own parallel Claude call.
Show the answer and explanation
Answer
C. Send each report to Claude as a PDF document, so every page arrives as an image along with its text.
How to read this question
- Situation. Defects written in the notes are all captured. Almost every defect visible only as a color band or a hand-drawn circle is missing. The script sends Claude only the text extracted from each PDF.
- What is asked. Where in the end-to-end pipeline the failure arises, and which change fixes it at that stage. One answer.
- Option by option.
- A. The script passes only plain text, so the charts and photos are not in the request at all. Telling the model what to look for cannot help it see content it never receives (reasoning from the scenario).
- B. OCR adds whatever text the images contain, such as axis labels. But the hot spots are color bands and the defects are unlabeled hand-drawn circles. The evidence is visual, so it still never reaches Claude (reasoning from the scenario).
- C (correct). The failure sits at the input stage, before Claude runs. With PDF support, each page is converted to an image and its text is supplied alongside, so Claude can see the color bands and circled defects. The trade-off is more input tokens per page (text cost plus image cost), and that cost is justified because the missing defects are the whole problem.
- D. Splitting the text and running parallel calls changes how the text is processed, but text extraction has already dropped the visual content (reasoning from the scenario).
- Rule to remember. When outputs miss information of one kind, check what the input stage delivers to the model before tuning the prompt or adding processing steps; nothing downstream can recover content dropped on the way in.
Sources (Claude Docs, quotes checked Oct 28, 2026)
- PDF support:
- "The system converts each page of the document into an image."
- "The text from each page is extracted and provided alongside each page's image."
- "This allows users to ask for insights on visual elements of a PDF, such as charts, diagrams, and other non-textual content."
- Cost of the trade-off: "Each page typically uses 1,500-3,000 tokens per page depending on content density." and "Because each page is converted into an image, the same image-based cost calculations are applied."
More practice
Questions
Where does this question come from?
It's an original sample question that CERTain·exam wrote based on Anthropic's public documentation and the public exam guide. Exam content is confidential under Anthropic's candidate agreement, and we do not use it.
What does CCAR-P Domain 1 cover?
According to the exam guide effective until December 6, 2026, Solution Design & Architecture (17%) covers translating business problems into Claude-based solutions, designing end-to-end architectures (input, processing, output and feedback loops), selecting architectural patterns (workflow, agentic, augmented LLM), multi-agent systems and orchestration, decomposition of complex problems, and aligning solutions to business value pillars. From December 7, 2026, the domain weighs 18%, and its objectives are rewritten. See what changes on December 7, 2026.
First, sign in with your email.