16 Governance, Safety & Stakeholders Practice Questions & Answers

1. 1. A hospital wants to use Claude via API to summarize clinical notes that contain protected health information (PHI). As the architect, which requirement must be satisfied before any PHI is processed?

   • A.Execute a Business Associate Agreement (BAA) with the API provider and use a HIPAA-eligible service configuration✓ Answer
   • B.Obtain individual written consent from every patient for AI processing
   • C.Fully de-identify all records, since PHI can never be sent to an API under any arrangement
   • D.Register the AI system with the Department of Health and Human Services before deployment

   Under the HIPAA Privacy Rule, a covered entity may disclose PHI to a vendor that processes it on its behalf only if that vendor is bound as a business associate through a BAA, which contractually obligates the vendor to safeguard the PHI. Individual patient consent is not required for treatment and operations uses, de-identification is one option but not mandatory once a BAA and eligible configuration are in place, and there is no HHS pre-deployment registration requirement.

   Source: HIPAA Privacy Rule, 45 CFR 164.502(e) — business associate contracts (BAA requirement)Report a problem with this question

2. 2. You are designing a RAG system whose knowledge base includes EU customers' personal data. Which design best supports GDPR's right to erasure?

   • A.Continue storing the data but filter it out of responses shown to the requesting user
   • B.Keep personal data only in queryable stores (documents and vector index entries) with a deletion pipeline that can locate and purge an individual's records on request✓ Answer
   • C.Explain to customers that data cannot be removed once it has been embedded
   • D.Anonymize all stored personal data ten years after collection

   GDPR Article 17 gives data subjects the right to have their personal data erased without undue delay, so the architecture must keep personal data in stores where individual records can be found and actually deleted — such as source documents and their vector index entries — rather than baked into artifacts that cannot be selectively purged. Merely hiding data from one user or deferring anonymization for years does not satisfy erasure, and claiming deletion is impossible is a design failure, not a legal defense.

   Source: GDPR Article 17 — Right to erasure ('right to be forgotten')Report a problem with this question

3. 3. A U.S. federal agency wants to deploy a Claude-based assistant for internal use with government data. Which compliance requirement is most directly relevant to selecting the hosting environment?

   • A.Any commercial API endpoint is acceptable as long as TLS encryption is used in transit
   • B.The service must run in a FedRAMP-authorized cloud environment at the impact level required for the data✓ Answer
   • C.The vendor's own self-attestation of security controls is sufficient without independent assessment
   • D.The agency must sign a HIPAA Business Associate Agreement with the model provider

   Federal agencies are required to use cloud services that hold a FedRAMP authorization at the appropriate impact level (Low, Moderate, or High) because FedRAMP standardizes security assessment and continuous monitoring for cloud products used by the government. TLS alone addresses only transport security, a HIPAA BAA governs health data rather than federal hosting, and FedRAMP specifically requires assessment by an accredited third-party assessment organization rather than vendor self-attestation.

   Source: FedRAMP Authorization Act, 44 U.S.C. § 3609 — federal agency use of FedRAMP-authorized cloud services with 3PAO assessmentReport a problem with this question

4. 4. An agent built on Claude reads content from external websites and has tools that can send emails on the user's behalf. Which control is the strongest defense against prompt injection embedded in retrieved web content?

   • A.Filter out all uppercase text from retrieved pages before passing them to the model
   • B.Scope tools to least privilege, treat retrieved content as untrusted data, and require human confirmation before consequential actions like sending email✓ Answer
   • C.Add a line to the system prompt telling the model to ignore any malicious instructions it encounters
   • D.Fine-tune the model on examples of safe email behavior

   Because no prompt-level instruction can guarantee the model will never be influenced by adversarial text, robust agent design applies defense in depth: architectural controls that limit what a compromised turn can actually do — least-privilege tool scopes, treating external content as data rather than instructions, and a human approval gate before irreversible or consequential actions. System-prompt warnings and fine-tuning reduce but do not eliminate the risk, and superficial text filters do not address the attack vector at all.

   Source: Claude documentation — agent safety and prompt-injection mitigations (defense in depth, least-privilege tools, human confirmation)Report a problem with this question

5. 5. A legal research assistant occasionally cites court cases that do not exist. Which combination of mitigations most directly addresses this failure mode?

   • A.Increase the sampling temperature so the model explores more diverse answers
   • B.Ask the model to review and confirm its own answer in a second turn, with no other changes
   • C.Ground responses in retrieved verified sources with citations, explicitly allow the model to say it cannot find an answer, and require human verification before any filing✓ Answer
   • D.Raise the maximum output token limit so the model can explain its reasoning at length

   Hallucination is mitigated by grounding: retrieval ties every citation to a verifiable source document, giving the model explicit permission to express uncertainty removes the pressure to fabricate, and human review catches residual errors before they cause harm in a high-stakes legal context. Raising temperature increases variability rather than accuracy, longer outputs do not improve factual grounding, and unaided self-review often fails because the model lacks any external reference to check against.

   Source: Claude documentation — reducing hallucinations (grounding with sources, allowing uncertainty, verification)Report a problem with this question

6. 6. You are architecting an insurance claims workflow where Claude drafts decisions. Where is human-in-the-loop review most critical?

   • A.Before irreversible or consequential decisions such as claim denials, with low-confidence cases routed to human adjusters✓ Answer
   • B.Only during initial testing, after which review can be fully removed
   • C.Reviewing every model input before it is submitted, regardless of content
   • D.Sampling 1% of outputs for audit after the actions have already been executed

   Human oversight delivers the most risk reduction when it gates the point of consequence: an adverse decision like a claim denial directly affects a consumer, so a human must review it before it takes effect, and confidence-based routing concentrates scarce reviewer attention on the cases most likely to be wrong. Reviewing all inputs adds cost without catching output errors, removing review after testing ignores drift, and post-hoc sampling audits detect harm only after it has occurred.

   Source: Anthropic Usage Policy — human-in-the-loop requirements for high-risk use cases affecting consumer outcomesReport a problem with this question

7. 7. A company deploys a Claude-based assistant to screen job application materials. Which approach best validates fairness before and after deployment?

   • A.Test only with synthetic candidate profiles generated by the same model
   • B.Remove candidate names from documents and consider the bias problem solved
   • C.Run evaluations on representative datasets and measure outcome disparities across demographic groups, repeating the analysis in production✓ Answer
   • D.Rely on the model vendor's general statement that the model was trained to be unbiased

   Fairness must be demonstrated empirically: measuring selection-outcome disparities across demographic groups on representative data is the mechanism behind disparate-impact analysis, and repeating it in production catches drift that pre-launch testing cannot. Removing names is insufficient because proxy variables (schools, zip codes, activities) still correlate with protected characteristics, vendor assurances are not evidence for your specific use case, and synthetic data from the same model inherits its own biases.

   Source: EEOC Uniform Guidelines on Employee Selection Procedures, 29 CFR Part 1607 — adverse (disparate) impact analysisReport a problem with this question

8. 8. A company launches a Claude-powered customer support chatbot. What does the transparency principle require regarding end users?

   • A.Disclose to users that they are interacting with an AI system rather than a human✓ Answer
   • B.No disclosure is needed as long as answer quality matches a human agent
   • C.Disclose that the system is AI only if a customer directly asks
   • D.Log all conversations for model improvement without informing users

   Proactive disclosure is required because users cannot make informed decisions about what to share or how much to rely on answers if they mistakenly believe they are talking to a human; Anthropic's Usage Policy requires disclosing AI interaction in consumer-facing deployments, and passing an AI off as a human can also constitute a deceptive practice under consumer-protection law. Waiting to be asked, pointing to output quality, or silent logging all fail this affirmative duty.

   Source: Anthropic Usage Policy — disclosure requirement that users know they are interacting with an AI; FTC Act Section 5 (deceptive practices)Report a problem with this question

9. 9. During structured discovery for a new Claude-based solution, what should the architect establish first?

   • A.The service-level agreement terms for the production system
   • B.A working prototype to show stakeholders something tangible immediately
   • C.The business problem, measurable success criteria, and constraints, before proposing any technical design✓ Answer
   • D.The model size and context window the solution will use

   Requirements precede solution design: the business problem and its measurable success criteria determine everything downstream — what to evaluate, which model and architecture fit, what risks matter, and what an SLA should even promise — so choosing technology or committing to SLAs first optimizes against unknowns. Prototyping is valuable, but only after the team knows what outcome the prototype must demonstrate.

   Source: CCAR-P Exam Guide, Domain 6 — structured discovery and requirements gathering (problem definition before solution design)Report a problem with this question

10. 10. Which single question is most useful for determining whether an LLM is an appropriate fit for a proposed workflow?

    • A.Is the engineering team enthusiastic about working with generative AI?
    • B.Can the model be hosted on the company's own GPU hardware?
    • C.Are our competitors already using LLMs for a similar workflow?
    • D.What is the cost of an incorrect output, and can errors be detected and tolerated in this workflow?✓ Answer

    LLMs are probabilistic systems that will sometimes produce wrong outputs, so fitness for a use case hinges on the consequence of error and the feasibility of catching it: workflows where mistakes are cheap, detectable, or reviewable are good fits, while workflows with silent, high-cost failures require redesign or human gates. Competitor behavior, team enthusiasm, and hosting preferences say nothing about whether the failure modes are survivable.

    Source: CCAR-P Exam Guide, Domain 5 — LLM risk, limitation, and failure-mode identification (risk-based fit assessment)Report a problem with this question

11. 11. The CEO pressures you to skip the evaluation phase and launch a Claude-based product two months early. As the architect, what is the most professional response?

    • A.Quietly keep the original timeline and hope the delay goes unnoticed
    • B.Present quantified tradeoffs — pilot error rates, the projected cost of production failures, and a minimal evaluation plan that still protects the launch date✓ Answer
    • C.Comply fully and skip evaluations, since the CEO owns the business risk
    • D.Refuse the request and escalate the disagreement directly to the board of directors

    An architect's duty in stakeholder communication is to make risk legible in business terms and offer options, because executives can only make sound tradeoff decisions when the cost of skipping evaluation is quantified against the value of launching early; proposing a right-sized evaluation plan converts a yes/no standoff into a negotiable scope. Silent compliance hides known risk, covert delay destroys trust, and immediate escalation skips the collaborative step that usually resolves the conflict.

    Source: CCAR-P Exam Guide, Domain 6 — communicating architectural tradeoffs and risk to executive stakeholdersReport a problem with this question

12. 12. When defining a latency SLA for a streaming LLM application, which approach is most appropriate?

    • A.Guarantee zero downtime and identical latency for every request
    • B.Guarantee only the mean (average) response latency across all requests
    • C.Promise the same end-to-end latency regardless of prompt and output length
    • D.Set percentile targets (e.g., p95/p99) and use time-to-first-token as a key metric for streaming responses✓ Answer

    LLM latency varies widely with output length and load, so averages hide the tail experience that users actually feel; percentile targets bound the worst realistic cases, and time-to-first-token captures perceived responsiveness in streaming interfaces where total generation time depends on how much text is produced. Guaranteeing zero downtime or uniform latency is unachievable and commits the team to an SLA it will inevitably breach.

    Source: CCAR-P Exam Guide, Domain 6 — SLA definition; SRE latency SLO convention (percentile-based targets, TTFT for streaming)Report a problem with this question

13. 13. Your team is handing off a production Claude application to an operations team. Which documentation is most critical to include in the handoff package?

    • A.The source code alone, since well-written code is self-documenting
    • B.A runbook covering monitoring alerts, escalation paths, rollback procedures, prompt/configuration versions, and evaluation baselines for detecting regressions✓ Answer
    • C.The original marketing deck used to pitch the project internally
    • D.Complete chat transcripts of every experiment the development team ran

    An operations team's job is to keep the system healthy and respond to incidents without the original builders present, which requires exactly the operational knowledge a runbook encodes: what the alerts mean, who to escalate to, how to roll back safely, which prompt and configuration versions are live, and what evaluation baselines define 'normal' so regressions are detectable. Source code and raw experiment transcripts do not tell an on-call engineer what to do at 3 a.m., and LLM applications in particular need prompt versioning and eval baselines because behavior can shift without any code change.

    Source: CCAR-P Exam Guide, Domain 6 — documentation and operational handoff (runbooks, versioning, eval baselines)Report a problem with this question

14. 14. A Claude-based application has been deployed to production. According to lifecycle best practices, what is the most important ongoing activity?

    • A.Freeze the system entirely so that nothing can change after launch
    • B.Disable logging in production to reduce storage costs
    • C.Continuously monitor output quality for drift and re-run the evaluation suite whenever prompts, models, or data sources change✓ Answer
    • D.Retrain the underlying model weights on a weekly schedule

    Deployment is not the end of the lifecycle: LLM application behavior can shift when user inputs drift, knowledge sources change, or the underlying model is updated, so continuous monitoring plus re-running evaluations on every meaningful change is the mechanism that keeps quality claims true over time. Freezing the system prevents necessary fixes, application teams generally cannot retrain a hosted model's weights, and disabling logging removes the very observability that monitoring depends on.

    Source: CCAR-P Exam Guide, Domain 6 — lifecycle phases (post-deployment monitoring and evaluation regression)Report a problem with this question

15. 15. An engineering lead wants every developer on the team to get consistent behavior from Claude Code — same project conventions, same tool permissions. What is the recommended approach?

    • A.Explain the conventions verbally during onboarding and rely on developers to remember them
    • B.Commit a CLAUDE.md file and shared project settings (.claude/settings.json) to the repository so every teammate's sessions inherit the same instructions and permissions✓ Answer
    • C.Have each developer maintain their own private configuration in their home directory
    • D.Distribute a PDF of best practices by email and ask developers to follow it manually

    Claude Code automatically loads a project's CLAUDE.md into context and applies checked-in project settings, so committing both to version control is the mechanism that makes conventions and tool permissions apply identically and automatically to every teammate's sessions — reviewed, versioned, and updated like any other code. Per-developer home-directory configs diverge silently, and PDFs or verbal onboarding depend on human memory rather than being enforced by the tool.

    Source: Claude Code documentation — project memory (CLAUDE.md) and shared project settings (.claude/settings.json) checked into version controlReport a problem with this question

16. 16. A platform team must guarantee that no developer's Claude Code session can ever execute certain dangerous commands, regardless of individual preferences. Which mechanism achieves this?

    • A.Add deny rules to each developer's personal user settings file
    • B.Rely on the model itself to refuse dangerous commands when asked
    • C.Ask each developer to be careful and avoid running those commands
    • D.Deploy managed policy settings with permission deny rules, which take precedence over and cannot be overridden by user or project settings✓ Answer

    Claude Code resolves settings in a strict precedence order in which enterprise managed policy settings sit at the top, so deny rules placed there are enforced by the harness itself and cannot be relaxed by any user-level or project-level configuration — making them the only option that provides a guarantee rather than a convention. Personal settings files can be edited by their owners, and both developer discipline and model refusals are probabilistic safeguards, not enforced controls.

    Source: Claude Code documentation — settings precedence (enterprise managed policy settings override user and project settings; permission deny rules)Report a problem with this question

Practice questions based on the official Claude Certified Architect – Professional (CCAR-P) exam guide and Anthropic's public documentation. This is an independent study tool, not affiliated with or endorsed by Anthropic, and does not grant certification. The real exam is 63 questions, 120 minutes, passing at a scaled 720/1000, delivered via Pearson VUE ($175). Official certification page →