CCAR-P Practice Exam 2 — 20 Questions

Question 1

A company is integrating Claude into a high-stakes automated decision-making system. Which TWO practices should be implemented to align with Anthropic’s Responsible AI principles?

Accepted Answer

Implement human-in-the-loop oversight for model outputs.. Integrating LLMs into high-stakes environments requires a focus on human-in-the-loop (HITL) oversight and continuous monitoring. These practices help manage hallucinations and maintain accountability. Without these safeguards, automated systems can produce biased or incorrect outcomes without human correction. Implementing these protocols ensures that AI governance remains aligned with human values and organizational safety goals, reducing the risk of unintended consequences in critical decision-making processes.

Answer

Allow the model to finalize decisions without human review.

Answer

Only use the most advanced model available for all tasks.

Answer

Disable all system logs to preserve user privacy.

Question 2

Your team is deploying an application that uses PII in prompts. A stakeholder asks how you are mitigating the risk of data leakage. How do you respond?

Accepted Answer

Explain the use of PII masking and data sanitization pipelines.. The correct response involves explaining a combination of data sanitization, prompt masking, and secure infrastructure design. This is critical because PII handling is a high-liability area. By detailing a defense-in-depth strategy, the architect provides the necessary assurance that privacy is treated with the highest priority, which is vital for maintaining organizational compliance and preventing severe legal or reputational damage during the application's lifecycle.

Answer

Assure them that the model is smart enough to handle PII.

Answer

Tell them the model does not store any data anyway.

Answer

Suggest moving the project to an on-premise LLM to ensure security.

Question 3

Refer to the exhibit. The user is attempting to trick the model into revealing sensitive information by claiming a high-clearance role. This is an example of which security threat, and how does the 'system' prompt help mitigate it?

Accepted Answer

Prompt injection (jailbreaking); the system prompt establishes a higher-priority context.. Social engineering and role-playing are common techniques used in prompt injection attacks to bypass security constraints. The system prompt is a powerful governance tool because it sets the foundational rules and persona for the model, which are prioritized by Claude's reasoning engine over conflicting instructions found in the user messages.

Answer

Man-in-the-middle attack; the system prompt encrypts the secret code name.

Answer

Denial of Service (DoS); the system prompt limits the tokens used to hide the secret.

Answer

Data poisoning; the system prompt cleans the training data in real-time.

Question 4

An agentic system is struggling with 'context fragmentation' over long-running sessions. What is the most effective architectural solution?

Accepted Answer

Implement a hierarchical memory system with summarization.. Context fragmentation happens when the history becomes too large or disorganized. A 'summarization agent' or a 'memory manager' that periodically compresses the conversation history into a concise summary is the standard solution. This preserves core context while discarding transient details, ensuring the main agent remains focused on the long-term goal rather than getting lost in thousands of lines of previous chat history.

Answer

Increase the context window size of the model indefinitely.

Answer

Force the user to clear their history periodically.

Answer

Only use the most recent 10 messages of the history.

Question 5

Refer to the exhibit. An agentic workflow encounters an error immediately after this message is generated by Claude. No further messages are sent to the API. What is the most likely cause of the failure in the orchestration logic?

Accepted Answer

The orchestrator failed to send a tool_result message. The Anthropic Messages API requires a specific sequence for tool use. After the assistant generates a 'tool_use' block, the orchestration layer must execute the tool and return a 'tool_result' message to the model before the assistant can continue its turn. Failing to provide this required response breaks the synchronous nature of the tool-calling conversation flow.

Answer

The input parameters do not match the JSON schema

Answer

The tool_use ID is missing the required prefix

Answer

The assistant content block lacks a stop_reason

Question 6

Which THREE factors should be prioritized when selecting an Anthropic model for a production-grade application?

Accepted Answer

Latency requirements of the specific user task.. Balancing performance, latency, and cost is fundamental to operational enablement. A professional architect must consider the specific requirements of the task—whether high-level reasoning or rapid response—to ensure that the chosen model delivers value efficiently. These factors dictate the system's scalability and overall budget, making them the primary drivers for architectural decisions when building and maintaining reliable AI-powered solutions in a corporate environment.

Answer

The total number of parameters in the model.

Answer

Whether the model is open-source or proprietary.

Question 7

A development team wants to optimize the latency of their prompt engineering workflow using Claude. They currently run evaluations sequentially. Which approach best improves iteration speed?

Accepted Answer

Implement a distributed asynchronous evaluation framework for concurrent prompt execution.. Parallelizing evaluation pipelines allows developers to test multiple prompt variations concurrently, significantly reducing feedback loops. In the context of LLM development, bottlenecking often occurs during the testing phase where prompt sensitivity to small changes requires broad regression coverage. By integrating asynchronous evaluation frameworks into CI/CD, teams can validate changes rapidly without manual overhead, ensuring that prompt performance remains consistent across diverse input datasets.

Answer

Switch to a smaller model version for all initial prompt testing.

Answer

Reduce the number of test cases to ensure the evaluation suite completes quickly.

Answer

Cache all user prompts to avoid re-sending identical requests to the API.

Question 8

Which TWO practices best improve the maintainability of large-scale prompt libraries? (Choose TWO)

Accepted Answer

Store prompts in external configuration files or a dedicated prompt management system.. Managing large prompt libraries requires modularity and version control to ensure consistency. By treating prompts as code and decoupling them from application logic, teams can implement standardized testing and deployment workflows. This separation allows developers to iterate on prompt performance independently of the software release cycle, reducing the risk of regressions and enabling rapid experimentation within established operational guardrails.

Answer

Hardcode all prompt versions directly into the application source code.

Answer

Combine prompts and business logic into monolithic classes for better encapsulation.

Answer

Avoid using variables in prompts to ensure absolute predictability.

Question 9

A developer needs to log all prompt-response pairs for compliance. What is the most reliable way to implement this without impacting application performance?

Accepted Answer

Use an asynchronous message queue to offload log processing.. Asynchronous logging is the key to maintaining application performance while ensuring full auditability. By offloading the logging process to a background task, the main request thread is not blocked, ensuring that users experience minimal latency. This pattern is crucial for compliance-heavy environments where audit trails are non-negotiable but system performance must remain highly responsive to meet end-user demands.

Answer

Write logs to a local file system before sending the response to the user.

Answer

Request the user to manually send a copy of the interaction for audit.

Answer

Only log errors instead of every prompt-response pair.

Question 10

You are building a customer-support agent with Claude that must call a lookup_order tool, then a refund_order tool that depends on the order's status. During testing, the agent sometimes calls refund_order before the lookup_order result returns. Which change to your orchestration loop best enforces the required ordering?

Accepted Answer

Expose only lookup_order until its result is returned, then add refund_order to the tools array for the subsequent request.. Deterministic ordering is achieved by controlling what the model can do at each step, not by prompting harder. Withholding the dependent tool until the prerequisite result is in the conversation removes the failure mode entirely, because the model cannot select a tool absent from the current tools array. This makes the dependency a property of the orchestration state machine rather than of model compliance.

Answer

Increase the max_tokens value so the model has more room to reason about the correct order before calling tools.

Answer

Include both tools in the same request's tools array and let the model choose the call order.

Answer

Set tool_choice to force a tool call on every turn so the agent never stalls between lookup and refund.

Question 11

A support agent built on Claude must decide whether a customer request is a billing issue, a technical issue, or a general inquiry before routing it. The categories are fixed and mutually exclusive, and the routing decision must be fast and cheap. Which approach is most appropriate?

Accepted Answer

Use a single Claude call with a constrained output format such as a tool definition or structured JSON that returns exactly one category label.. Fixed, mutually exclusive categories with strict latency and cost goals are best handled by a single Claude call that constrains output to one of the allowed labels. Using a tool definition or JSON schema with an enum removes ambiguity and avoids the overhead of multi-agent orchestration or a separate fine-tuned model.

Answer

Fine-tune a separate model on historical tickets and deploy it alongside the Claude agent for classification.

Answer

Give the agent access to all internal tools and let it infer the category from whichever tool it chooses to call.

Answer

Build a multi-agent pipeline where a classifier agent debates a router agent until they agree on a category.

Question 12

An agent uses a retrieval tool that returns the top 20 chunks for any query. In production, the agent frequently cites irrelevant chunks and sometimes misses the correct answer even when it is present in the corpus. The corpus contains documents with overlapping terminology. Which architectural change most improves answer grounding without increasing the number of retrieved chunks?

Accepted Answer

Add a re-ranking stage that scores the retrieved chunks against the query and passes only the highest-scoring subset to the model.. The retrieval step is high-recall but low-precision, which is why irrelevant chunks are cited and correct ones are missed amid overlapping terminology. Adding a re-ranking stage that scores candidates against the query and passes only the top subset improves precision without retrieving more chunks, directly improving grounding and citation quality.

Answer

Embed the entire corpus into the system prompt so the model always has full context available.

Answer

Increase the retrieval count to 50 so the correct chunk is more likely to be included somewhere in the context.

Answer

Lower the model's temperature to zero so it sticks more closely to the retrieved text.

Question 13

An orchestration agent runs a nightly workflow that fans out to 12 subagents, each calling a partner REST API. Partner calls intermittently return HTTP 429 with a Retry-After header. The orchestrator currently retries immediately in a tight loop, causing cascading 429s and duplicate side effects on the partner systems. Which architectural change best addresses both the throttling and the duplicate side effects?

Accepted Answer

Introduce a shared token-bucket rate limiter in front of the partner calls plus idempotency keys on every mutating request, with subagents respecting Retry-After.. Bursty fan-out against a rate-limited partner needs two coordinated controls: a shared limiter so the aggregate request rate stays under the partner's ceiling, and idempotency keys so a retried mutating call cannot create a second side effect. Honoring Retry-After aligns retry timing with the partner's guidance. Token or context changes operate on model text, not on outbound traffic, and serializing the workload removes parallelism without addressing throttling or duplication.

Answer

Reduce the fan-out to a single subagent that processes all 12 partners sequentially with no retry logic at all.

Answer

Switch every subagent to a larger context window so each can hold the full history of prior 429 responses and learn to avoid them.

Answer

Increase the orchestrator's max_tokens so the model can reason longer about each 429 before deciding to retry.

Question 14

You are architecting a Claude-based agent for a regulated financial client that performs long-running portfolio rebalancing workflows. A compliance requirement mandates that no single trade instruction may be executed unless it is cryptographically traceable to the exact model-generated intent that produced it. The agent uses the Messages API with tool use, and several downstream services consume tool calls asynchronously. Which architectural mechanism best satisfies this requirement while preserving agent autonomy?

Accepted Answer

Persist every assistant turn containing a tool_use block, its tool_use id, and the matching tool_result, forming an immutable audit chain keyed by the tool_use id.. Binding each tool call to the platform-generated tool_use id and persisting the surrounding assistant turn creates a deterministic chain from model intent to executed instruction. Because the id appears in both the tool_use block and the tool_result, downstream services and auditors can reconcile them without trusting model-authored text. Caching, summaries, or self-hashes do not establish that binding.

Answer

Require the model to emit a SHA-256 hash of its own reasoning text inside the tool input, and verify that hash in the downstream trade service before execution.

Answer

Route all trades through a single orchestrator agent that logs a natural-language summary of each decision to an append-only ledger after the trade is confirmed.

Answer

Enable prompt caching on the system prompt so that every trade instruction inherits a stable cache key that can be presented to auditors as proof of origin.

Question 15

Your team operates a Claude agent that triages inbound customer support tickets. After several weeks in production, the agent begins approving refunds above the policy ceiling and citing outdated policy text. Investigation shows the system prompt embeds a policy document that was updated three weeks ago, but the deployed prompt was never regenerated. Which architectural practice most directly prevents this class of failure?

Accepted Answer

Move policy content out of the system prompt and retrieve the current policy at runtime via a versioned tool or retrieval step, recording the version used in each decision.. Stale behavior traced to a frozen prompt is best solved by externalizing the mutable knowledge and fetching it at decision time from a versioned source. That way policy updates propagate immediately, and logging the retrieved version makes each action auditable. Compression, randomness, or peer review all leave the embedded copy as the de facto source of truth.

Answer

Add a second Claude agent that reviews the first agent's refund decisions and overrides any that appear inconsistent with its own recollection of company policy.

Answer

Increase the model's temperature slightly so the agent explores alternative interpretations of the policy text it has memorized.

Answer

Shorten the system prompt by summarizing the policy into a compact bullet list, which reduces the chance that outdated sections remain embedded.

Question 16

A retail company is building a governance program for a customer-facing Claude agent that can issue refunds and update order records. The risk committee wants controls that limit the blast radius of a compromised or misbehaving agent. (Choose two.)

Accepted Answer

Require human approval for refunds above a defined monetary threshold before the action is executed.. Blast-radius reduction requires preventive controls that bound what the agent can execute. Least-privilege tool scopes with per-action authorization shrink the callable action set, and threshold-based human approval stops high-value refunds before they happen. Logging, policy publication, and periodic transcript review are detective or administrative measures that document or discourage misuse but do not constrain the agent's real-time capabilities, so they fail the risk committee's objective.

Answer

Enable verbose debug logging of every prompt and response for later forensic review.

Answer

Schedule quarterly reviews of the agent's conversation transcripts to identify emerging misuse patterns.

Answer

Publish an acceptable use policy that prohibits employees from manipulating the agent.

Question 17

A fintech company's risk committee is operationalizing a governance program for Claude-powered customer support agents. They must demonstrate to regulators that model behavior changes are tracked, attributable, and reversible. Which TWO practices best satisfy this requirement? (Choose two.)

Accepted Answer

Run a fixed regression suite against each candidate model version and archive the scored results before promoting it.. Attributable, reversible model change management requires two things working together: a frozen, dated artifact in production and comparable evidence captured before promotion. Pinning to dated model identifiers supplies the frozen artifact and the rollback target, while a fixed regression suite run against each candidate supplies the before-and-after evidence. Together they let the risk committee answer what changed, who approved it, and how to undo it.

Answer

Route every request through a gateway that logs the full request and response payloads with the model identifier attached.

Answer

Enable prompt caching on all production requests so that previously validated prompts are reused verbatim.

Answer

Store the system prompt in a version-controlled repository and require pull-request review before it is merged.

Question 18

A multinational bank deploys Claude to draft internal policy summaries for staff in the EU and Singapore. Legal requires that personal data embedded in employee questions never leave its region of origin, but the bank wants a single application codebase. Which architecture most directly enforces the residency requirement?

Accepted Answer

Configure regional API endpoints so that EU traffic is served within the EU and Singapore traffic within its region, sharing only stateless application code.. Data residency is a geographic constraint on where processing happens, so the control must shape the request path rather than the contract or the retention window. Routing each jurisdiction to an in-region API endpoint keeps personal data inside its required boundary while a shared, stateless codebase avoids duplicated engineering effort. Redaction and retention settings reduce risk but do not guarantee the data never leaves the region.

Answer

Enable zero data retention on the account so that no request or response content is stored by the provider.

Answer

Deploy the application in one region and rely on the model provider's contractual data processing addendum to cover cross-border transfer.

Answer

Apply client-side redaction of names and identifiers before the request is sent, then send all traffic to a single global endpoint.

Question 19

A hospital network is drafting its AI risk register for a Claude-based discharge-summary assistant. The governance lead wants entries that describe residual risk after existing controls are applied, and that can be assigned an owner and a review cadence. Which TWO characteristics must each risk register entry have to meet this standard? (Choose two.)

Accepted Answer

A residual risk rating that reflects the effect of the controls already in place.. An operational risk register entry must state the exposure that remains after controls and must have someone accountable for watching it on a schedule. Residual rating captures what is actually at stake post-mitigation, while owner and review interval ensure the entry is revisited and acted upon as the assistant and its clinical context change.

Answer

A verbatim copy of the vendor's model card and system prompt.

Answer

A list of every employee who has ever accessed the assistant.

Answer

A projected cost saving attributed to deploying the assistant.

Question 20

A stakeholder asks for a change in the model's tone to be 'more professional' for customer support. How should you translate this request into technical requirements?

Accepted Answer

Create a clear system prompt with guidelines for tone, style, and vocabulary.. Translating subjective feedback like 'professional tone' into concrete system instructions is a core task for an architect. By creating a specific system prompt and testing it against a set of representative inputs, you define what 'professional' means for the system. This provides a clear, verifiable standard that ensures consistent communication across all customer support interactions, improving both the quality of service and the reliability of the AI implementation.

Answer

Tell the stakeholder that tone is subjective and suggest we keep the default setting.

Answer

Change the model version to one that is specifically advertised as 'professional'.

Answer

Ask the stakeholder to rewrite the model's responses themselves to ensure it is correct.