Timed Mode Diagnostic Set – Claude Certified Architect Professional CCAR-P

1. Question

A global retailer uses Claude to review inbound freight quotes before selecting carriers for weekly shipments. The system must first extract shipment costs and service terms, then validate the results against procurement policies, and finally generate a recommendation for the transportation team. Because the stages follow consistent rules and later stages depend on the output of earlier stages, the company needs a predictable and auditable architectural pattern.

Which architectural pattern best meets these requirements?

Use a fixed workflow where each step is a discrete, sequenced LLM call for extracting quote details, validating policy compliance, and generating the recommendation.
Use a single monolithic prompt that asks Claude to interpret the quotes and autonomously decide whether additional processing or external actions are required.
Use an autonomous agent that dynamically creates its own processing plan and selects tools based on each set of inbound freight quotes.
Use parallel LLM calls that independently extract, validate, and generate the recommendation at the same time without waiting for preceding results.

2. Question

A customer support platform uses Claude to classify short messages into predefined categories. The team has a labeled test set for comparing models in the same model family and needs a model strategy that balances accuracy, latency, and cost.

Which model strategy best meets these requirements?

Use the largest model in the model family for every message because maximum capability should take priority over latency and cost.
Use the fastest model after testing only a few sample messages, without performing ongoing evaluation.
Select a model based only on message length instead of measured classification accuracy.
Select the smallest model that satisfies the accuracy requirement, and continuously verify its performance through ongoing evaluations.

3. Question

A company uses Claude across multiple AI applications that need access to the same internal systems. The architecture team wants to reuse the same connections and avoid building a separate integration for every application.

Which integration architecture best satisfies these requirements?

Build a separate custom API integration between every AI application and each internal system.
Provide access to each internal system through an MCP server that any compatible application can reuse.
Embed all system connection logic directly into the Claude prompts used by each application.
Route every AI application through a single Claude conversation that maintains persistent access to all internal systems.

4. Question

A legal operations team is assessing the launch readiness of a Claude-powered contract review system that performs clause extraction. Stakeholders need evidence that the system improves review operations while maintaining acceptable quality, and must agree on the conditions required before production deployment.

Which evaluation approach best supports the launch decision?

Measure general model benchmarks, average response length, and user adoption after deployment without establishing acceptance criteria.
Track only technical metrics such as token consumption, API latency, and system availability because operational performance determines launch readiness.
Establish task-specific metrics linked to business outcomes and set agreed performance thresholds for launch approval.
Ask legal reviewers whether the generated results appear helpful and launch once most participants provide positive qualitative feedback.

5. Question

A company sends Claude conversation records to an external platform for quality monitoring. Because the records contain personal data, the architecture team must address a GDPR concern before the data leaves the system boundary.

Which architectural approach best addresses this concern?

Encrypt the complete records during transfer and retain all identifiable information indefinitely.
Send the complete records and redact personal data only when a reviewer requests access.
Provide a privacy notice, then transfer and retain the complete records without modifying the personal data.
Apply data minimization by redacting or pseudonymizing personal information before external transfer and retaining only the data required for quality monitoring.

6. Question

Select the correct answer from the drop-down list of options. Each correct selection is worth one point.

A request-handling component assigns a risk category to submitted content, normalizes the conversation metadata, and attaches the resulting attributes to the payload used by the application:

Claude evaluates the conversation context, system instructions, and available tool schemas to produce a structured tool_use content block containing an operation name and input parameters:

A response-handling component extracts content blocks from the Claude API result, verifies that structured data conforms to an application-defined schema, and converts valid data into typed application objects:

An application combines retrieved policy excerpts, a summarized conversation history, and the latest submitted message into the final message payload supplied to Claude:

Claude incorporates information contained in a tool_result content block and produces either a natural-language answer or an additional structured tool request:[CCARP-D-05]

3. Question

An online marketplace updates product availability, discounts, and pricing several times each day. A customer-support assistant powered by Claude must provide responses based on the latest marketplace records. Which approach will best meet these requirements?

Fine-tune Claude monthly using a snapshot of the marketplace records.
Connect Claude to a retrieval system that queries the current marketplace data for each request.
Store a summarized catalog in the system prompt and refresh it automatically once each week.
Generate approved responses for common product questions after each nightly catalog update and return the closest matching response.

4. Question

A development team finds that an AI assistant inconsistently applies formatting rules. Which prompt-engineering approaches would most effectively make the assistant follow the specified instructions more consistently? (Select TWO.)

Define a detailed professional role and domain expertise for the assistant at the beginning of the prompt.
Separate the prompt into clearly labeled sections and define which instruction takes precedence when requirements conflict.
Include a brief explanation of the business purpose behind each instruction.
Combine all requirements into uninterrupted prose to make the prompt more conversational.
Add representative examples that demonstrate the expected response for frequently mishandled situations.

5. Question

A retail company’s purchasing agent must coordinate shipment dates with an autonomous logistics agent managed by an external distributor. Both companies require the internal applications, tools, and databases to remain inaccessible to the other party. Which integration method most appropriately meets the stated constraints?

Replicate selected scheduling records into a common database that both agents can query without mediation.
Enable Agent-to-agent communication through a mutually supported protocol, allowing each agent to retain controlled access to its organization’s private systems.
Connect both agents to a shared integration gateway that directly exposes each organization’s internal scheduling tools.
Consolidate the agents into one shared service administered by an independent provider.

6. Question

A financial institution is preparing to validate a generative AI assistant before allowing it to support compliance-related workflows. Which evaluation components should the institution prioritize? (Select TWO.)

A latency test that measures response speed under increasing numbers of concurrent requests.
A reference dataset reviewed and labeled by compliance specialists, including routine cases and high-impact edge cases.
A collection of adversarial prompts that tests jailbreak attempts, prompt injection, unsafe requests, and policy enforcement.
A general-purpose benchmark that compares the assistant’s broad knowledge with other foundation models.
A cost comparison that identifies which foundation model produces the lowest average inference expense.