CCAR-P Practice Exam

This is a full-length, 63-item practice exam for CCAR-P. The questions are new (not reused from the domain pages) and distributed to match the blueprint. Sit it under exam conditions before you rely on the score.

Instructions

Time: 120 minutes. Set a timer and do not pause it.
Format: multiple-choice (select one) and multiple-response (select two); each stem states how many to select.
Scoring: the real exam is scaled 100–1000 with a pass at 720. There is no guessing penalty — answer everything. As a raw proxy, aim for ≥ 80% (≈ 50/63) before booking.
Method: read the whole stem, identify the binding constraint, eliminate the constraint-blind / over-engineered / prompt-as-enforcement / self-report / silent-failure / recall-only / aggregate-metric distractors, then choose.

Domain distribution

#	Domain	Weight	Items here
1	Solution Design & Architecture	17%	11
2	Claude Models, Prompting & Context Engineering	13%	8
3	Integration (incl. RAG)	19%	12
4	Evaluation, Testing & Optimization	16%	10
5	Governance, Safety & Risk Management	14%	9
6	Stakeholder Communication & Lifecycle Management	14%	9
7	Developer Productivity & Operational Enablement	7%	4
	Total	100%	63

Score interpretation

Raw score (of 63)	Percent	Reading
57–63	90–100%	Exam-ready; strong across all domains
50–56	80–89%	Likely pass; shore up your weakest domain
45–49	71–79%	Borderline; drill D3/D1/D4 and re-sit
38–44	60–70%	Not ready; systematic gaps — restudy by domain
< 38	< 60%	Restudy the full course before re-attempting

The real exam scales 100–1000 with a pass at 720; treat ≥ 80% raw (≈ 50/63) here as your go/no-go line, and confirm you have no single domain far below the rest.

Take the practice exam

Two ways to use the questions below: the interactive mode runs a timed sitting one question at a time and ends with your score, a per-domain breakdown and a full correction; the review mode underneath lists every question with its options one per line and the answer hidden until you ask for it.

Interactive mode

Take the practice exam

63 questions · one at a time · 120-minute countdown · results with per-domain breakdown and full correction at the end. Your progress is saved in this browser if you leave the page.

Time limit Shuffle question order

CCAR-PPractice exam

Question 1 of 63

--:--

All questions (review mode)

Options are listed one per line. The answer and explanation stay hidden until you click Show answer. Use the interactive mode above for a timed sitting.

Q1D1 · Solution Design and ArchitectureSelect one
A logistics firm wants to automate shipment-status replies: 500/hour, one internal API, p95 under 4 s, tight cost ceiling. Which pattern should the architect propose?
- A. A coordinator with five specialised subagents.
- B. An augmented-LLM / simple workflow (classify → status lookup → structured reply) because the steps are known and the SLA and cost are binding.
- C. An open-ended agentic loop with a large tool catalogue.
- D. Fine-tuning before building any pipeline.
Show answer
Answer: B.
Known steps plus binding latency/cost favour the simplest pattern. Multi-agent (A) and broad agentic loops (C) over-engineer; fine-tuning (D) is premature.
Q2D1 · Solution Design and ArchitectureSelect one
A German bank requires all data to stay in the EU and already runs on AWS. Which placement is BEST?
- A. Direct Anthropic API.
- B. Amazon Bedrock in an EU region, inheriting existing AWS IAM and residency.
- C. Google Vertex AI.
- D. Whichever model is newest.
Show answer
Answer: B.
EU residency plus an existing AWS footprint point to Bedrock EU. Direct API (A) risks residency; Vertex (C) ignores the AWS investment; novelty (D) can't override residency.
Q3D1 · Solution Design and ArchitectureSelect one
An agentic loop occasionally runs forever. Which is the CORRECT primary control plus backstop?
- A. Parse the model's text for the word 'finished'.
- B. Terminate on stop_reason (end_turn / no further tool_use), with an iteration cap only as a safety backstop.
- C. A fixed 5-iteration cap as the sole control.
- D. Lower temperature until it stops.
Show answer
Answer: B.
Control flow keys off stop_reason; a cap is a backstop, not the mechanism. Text parsing (A) is brittle; a cap alone (C) is an anti-pattern; temperature (D) doesn't govern termination.
Q4D1 · Solution Design and ArchitectureSelect two
Peak load is 25 req/s at ~10k input tokens each, and the team hits ITPM limits. Which TWO are sound?
- A. Shift latency-tolerant work to the Message Batches API and add backoff with jitter honouring retry-after.
- B. Retry immediately in a tight loop.
- C. Request a higher tier and/or route overflow to a second model.
- D. Return empty results as success on 429.
- E. Send everything to Opus 5.
Show answer
Answer: A and C.
Batch for tolerant work plus backoff, and raising the tier / spillover routing, are the correct capacity levers. Tight-loop retry (B) worsens it; swallowing 429s (D) is silent failure; Opus for all (E) raises cost without fixing limits.
Q5D1 · Solution Design and ArchitectureSelect one
What makes an end-to-end reference architecture complete at Professional level?
- A. Input and processing stages only.
- B. Input → processing → output → a feedback loop that routes production signals back into prompts/models/retrieval.
- C. A single model call.
- D. Only a load balancer and the model.
Show answer
Answer: B.
The feedback loop is mandatory; a design with no path from production signals to improvement is incomplete. The others omit it.
Q6D1 · Solution Design and ArchitectureSelect one
A stem gives a hard $8,000/month budget and a modest accuracy bar for high-volume classification. Which posture is BEST?
- A. Opus 5 with xhigh effort.
- B. Haiku 4.5 (or Sonnet 5) with prompt caching and few tool round-trips, treating cost/volume as the binding pillars.
- C. Fable 5.1 for maximum quality.
- D. A multi-agent pipeline.
Show answer
Answer: B.
The binding pillar is cost at volume with a modest quality need — the cheapest adequate model with caching. Opus/Fable (A, C) overspend; multi-agent (D) adds cost.
Q7D1 · Solution Design and ArchitectureSelect one
When is multi-agent orchestration genuinely justified?
- A. A three-step sequential pipeline sharing context.
- B. Independent parallel investigations with isolated context that fan in to a synthesis step.
- C. A single API lookup with a tight SLA.
- D. Always, to be safe.
Show answer
Answer: B.
Parallel, context-isolated subtasks with fan-in are the multi-agent signal. Sequential shared work (A) is a workflow; a single lookup (C) is augmented-LLM; 'always' (D) is over-engineering.
Q8D1 · Solution Design and ArchitectureSelect one
The primary model returns sustained 529s under a spike. Which fallback is BEST?
- A. Error out to all users until recovery.
- B. Retry with backoff; if still failing, fall back to a secondary model/provider and log the request as served in degraded mode.
- C. Serve stale cached answers as if fresh.
- D. Fail an active Fable 5.1 thinking session over to a much older model.
Show answer
Answer: B.
Retry-then-fallback with degraded-mode logging preserves availability and observability. Hard failure (A) is poor HA; stale-as-fresh (C) is silent failure; downgrading a Fable 5.1 session (D) drops thinking blocks.
Q9D1 · Solution Design and ArchitectureSelect one
The company lacks ops capacity, needs launch in a month, and the capability is commodity. Build or buy?
- A. Build a custom Agent SDK runtime and sandbox.
- B. Use managed agents (Anthropic hosts the loop/sandbox) to hit the deadline with minimal ops.
- C. Delay until an ML platform team is hired.
- D. Buy a product with no Claude support.
Show answer
Answer: B.
No ops capacity + tight deadline + commodity capability → managed/buy. Building (A) contradicts the constraints; delaying (C) misses the deadline; (D) abandons the requirement.
Q10D1 · Solution Design and ArchitectureSelect one
Why capture a design as an ADR?
- A. To meet a documentation quota.
- B. To record context, rejected alternatives with reasons, and the accepted trade-off so the decision is reviewable and revisitable.
- C. Because ADRs replace evaluation.
- D. Because agentic systems require them technically.
Show answer
Answer: B.
ADRs make reasoning and trade-offs explicit and reviewable. They are not a formality (A), don't replace evals (C), and aren't a technical prerequisite (D).
Q11D1 · Solution Design and ArchitectureSelect one
A stakeholder gives a vague goal and no numbers. What is the BEST first action?
- A. Start building immediately.
- B. Run discovery to define measurable per-segment success criteria and enumerate constraints before designing.
- C. Default to Opus 5.
- D. Assume a target and proceed.
Show answer
Answer: B.
Without criteria/constraints the design is unanchored; discovery surfaces the binding constraint. Building (A), defaulting to a model (C), or inventing a target (D) skip anchoring.
Q12D2 · Claude Models, Prompting and Context EngineeringSelect one
A pipeline runs everything on Opus 5 at 4× budget with acceptable quality. Best cost fix preserving quality?
- A. Move all traffic to Haiku 4.5.
- B. Cascade cheap-first (Haiku/Sonnet), escalate to Opus 5 on validation-check failure, and cache the stable prefix.
- C. Escalate on the model's self-reported confidence.
- D. Raise effort to xhigh everywhere.
Show answer
Answer: B.
Cheap-first with escalation on validation failure preserves hard-case quality; caching amortises the prefix. Blanket Haiku (A) sacrifices quality; self-report (C) is an anti-pattern; xhigh everywhere (D) raises cost.
Q13D2 · Claude Models, Prompting and Context EngineeringSelect one
After upgrading to Sonnet 5, requests setting budget_tokens fail with 400. What is the fix?
- A. Add retries.
- B. Use thinking: {"type":"adaptive"} with effort, since budget_tokens is removed on Sonnet 5.
- C. Downgrade to Haiku 4.5 permanently.
- D. Remove thinking entirely.
Show answer
Answer: B.
budget_tokens is removed on Sonnet 5 (returns 400); adaptive thinking + effort replaces it. Retries (A) can't fix a 400; Haiku (C) is not equivalent; removing thinking (D) discards needed reasoning.
Q14D2 · Claude Models, Prompting and Context EngineeringSelect one
A cache hit rate is near zero because the volatile user turn is placed before the system prompt and documents. What is the fix?
- A. Disable caching.
- B. Reorder to stable-first (system → tools → docs), then the user turn, marking the stable boundary with cache_control.
- C. Shorten the documents.
- D. Switch to Haiku.
Show answer
Answer: B.
Caching keys on a stable prefix; the volatile turn must come last. Disabling (A) forfeits savings; length (C) only matters for the minimum; model choice (D) is irrelevant.
Q15D2 · Claude Models, Prompting and Context EngineeringSelect one
In an active Fable 5.1 session the team must change the system instructions mid-run. Correct approach?
- A. Edit the original system field.
- B. Append a new role: "system" message and leave earlier turns untouched (append-only harness).
- C. Delete the earliest turns.
- D. Reorder messages so the new instruction is first.
Show answer
Answer: B.
Fable 5.1 thinking blocks are invalidated by editing/reordering/removing earlier turns, so mid-session changes are appended. The other options invalidate downstream thinking.
Q16D2 · Claude Models, Prompting and Context EngineeringSelect one
A hard rule 'no discount above 25%' must be guaranteed. Where does it belong?
- A. A strongly worded system-prompt sentence.
- B. A programmatic tool-permission hook / validation that rejects any discount over 25% before execution.
- C. A few-shot example.
- D. The thinking budget.
Show answer
Answer: B.
Hard rules require deterministic enforcement. Prompt wording (A) and few-shot (C) are prompt-as-enforcement; the thinking budget (D) is unrelated.
Q17D2 · Claude Models, Prompting and Context EngineeringSelect one
A workload mandates Zero Data Retention but the team wants Fable 5.1. Correct guidance?
- A. Use Fable 5.1; ZDR is universal.
- B. Fable 5.1 requires 30-day retention and is not ZDR-eligible; choose a ZDR-eligible model (Opus 5 / Sonnet 5).
- C. Disable your logging to get ZDR on Fable 5.1.
- D. Use Fable 5.1 and delete logs at 30 days.
Show answer
Answer: B.
Fable 5.1's 30-day retention disqualifies it under a ZDR mandate. ZDR isn't universal (A); your logging (C) doesn't change provider retention; 30-day deletion (D) is what ZDR forbids.
Q18D2 · Claude Models, Prompting and Context EngineeringSelect one
Which statement about current-model thinking config is correct?
- A. All current models require budget_tokens.
- B. Current models use adaptive thinking with effort (low/medium/high/xhigh); only Haiku 4.5 still uses budget_tokens and lacks effort.
- C. effort exists only on Haiku 4.5.
- D. xhigh is the default everywhere.
Show answer
Answer: B.
Adaptive thinking + effort is standard; Haiku 4.5 is the budget_tokens exception without effort. The others misstate the API.
Q19D2 · Claude Models, Prompting and Context EngineeringSelect one
Reusable capability blocks are pasted into every prompt, bloating context and breaking the cache prefix. Better pattern?
- A. Package them as Skills loaded progressively on demand, keeping the cache prefix lean.
- B. Duplicate them into each service prompt.
- C. Move them into the volatile user turn.
- D. Enlarge the context window.
Show answer
Answer: A.
Skills load capability on demand and keep the prefix stable. Duplication (B) caused the bloat; the volatile turn (C) worsens caching; a bigger window (D) doesn't fix cost/caching.
Q20D3 · Integration (incl. RAG)Select one
A knowledge assistant keeps giving the OLD answer, confidently, after a document was updated overnight. What to investigate FIRST?
- A. Rewrite the system prompt.
- B. Inspect what retrieval returned; the updated doc was likely not re-chunked/re-embedded/re-indexed, so retrieval serves stale vectors.
- C. Switch to Opus 5 xhigh.
- D. Add few-shot examples.
Show answer
Answer: B.
Confident-wrong-after-refresh points to stale retrieval/indexing; inspect retrieved chunks first. Prompt changes (A, D) and a bigger model (C) can't fix stale retrieval.
Q21D3 · Integration (incl. RAG)Select one
A support agent has 18 tools including delete_account it should never use. Correct remediation?
- A. Keep it but log every call.
- B. Remove the unneeded tools from the allowlist (least privilege) so it cannot invoke them.
- C. Add an 'are you sure?' prompt.
- D. Forbid it in the system prompt.
Show answer
Answer: B.
Least privilege removes the capability. Logging (A) and confirmation (C) leave it reachable; a prompt rule (D) is bypassable.
Q22D3 · Integration (incl. RAG)Select one
Users search by exact SKU and by description; dense-only retrieval misses many SKUs. Best fix?
- A. Increase k to 1000.
- B. Hybrid retrieval (BM25 + dense with rank fusion) so exact identifiers and semantic matches both rank.
- C. A bigger generation model.
- D. Remove metadata filters.
Show answer
Answer: B.
Dense is weak on exact tokens; sparse handles SKUs and hybrid covers both. Huge k (A) adds noise; a bigger model (C) doesn't change retrieval; removing filters (D) harms precision/security.
Q23D3 · Integration (incl. RAG)Select one
recall@10 is 0.94 but answers include facts not in the retrieved chunks. Which layer and fix?
- A. Retrieval; lower k.
- B. Generation/grounding; enforce answer-only-from-context, add citations, and rerank so the best passage is on top.
- C. Embeddings; change the model.
- D. Indexing; re-index everything.
Show answer
Answer: B.
High recall means retrieval works; unsupported facts are a grounding failure. The retrieval-side fixes (A, C, D) target a healthy layer.
Q24D3 · Integration (incl. RAG)Select one
A 3M-document corpus changes daily and answers must cite the exact clause. Best approach?
- A. Weekly fine-tuning on the corpus.
- B. RAG with hybrid retrieval, reranking, citations, and a scheduled refresh pipeline.
- C. Stuff the whole corpus into 1M context per query.
- D. Long context + fine-tuning.
Show answer
Answer: B.
Large, changing, citation-requiring corpora are canonical RAG. Fine-tuning (A) can't track daily facts; the corpus exceeds/overspends context and loses citations (C); (D) inherits both issues.
Q25D3 · Integration (incl. RAG)Select two
Contracts chunked at fixed 500 tokens produce answers citing the wrong sub-clause and losing context. Which TWO help most?
- A. Structural/document-aware chunking on clause/section boundaries.
- B. Parent–child chunking: match small children, return the larger parent clause.
- C. Dense-only retrieval.
- D. Higher temperature.
- E. Removing citations.
Show answer
Answer: A and B.
Structured docs need boundary-aware chunking and parent–child context. Dense-only (C), temperature (D), and removing citations (E) don't address chunking.
Q26D3 · Integration (incl. RAG)Select one
A remote MCP server over Streamable HTTP exposes powerful tools with no auth. What must be added?
- A. Nothing; MCP is safe by default.
- B. OAuth 2.1 on the remote MCP server plus per-user permission checks inside the tools.
- C. A longer prompt.
- D. A higher rate-limit tier.
Show answer
Answer: B.
Remote MCP requires OAuth 2.1 and per-user authorization. MCP isn't authenticated by default (A); prompts (C) and tiers (D) don't address authz.
Q27D3 · Integration (incl. RAG)Select one
After enabling backoff retries, some customers are charged twice. Correct fix?
- A. Disable retries.
- B. Add idempotency keys to the charge action so retries de-duplicate with no duplicate side effect.
- C. Lower temperature.
- D. Reconcile duplicates weekly.
Show answer
Answer: B.
Idempotency keys make retries safe. Disabling retries (A) harms resilience; temperature (C) is irrelevant; reconciling later (D) still double-charged.
Q28D3 · Integration (incl. RAG)Select one
Some questions require follow-up lookups combining multiple sources. Which design fits and its trade-off?
- A. Agentic RAG (model decides when/what to retrieve, issues follow-ups), at the cost of more latency/tokens.
- B. Stuff everything into context.
- C. Fine-tune on the multi-hop set.
- D. Remove reranking for speed.
Show answer
Answer: A.
Multi-hop queries justify agentic RAG; the trade-off is cost/latency. Stuffing (B) doesn't scale; fine-tuning (C) can't hold changing facts; removing reranking (D) hurts precision.
Q29D3 · Integration (incl. RAG)Select one
A capability will be reused across many agents/clients under a standard protocol. Best integration mechanism?
- A. A per-app CLI tool copied into each service.
- B. An MCP server (OAuth 2.1 if remote) for standard, reusable access.
- C. Agent-to-agent handoff only.
- D. Pasting the logic into every prompt.
Show answer
Answer: B.
Cross-client reuse under a standard protocol is MCP's purpose. Per-app tools (A) fragment; agent-to-agent (C) is for context isolation; prompt-pasting (D) is unmaintainable.
Q30D3 · Integration (incl. RAG)Select two
An agent loads all 40 tools and 30 docs each request: high cost, low cache hits, poor tool selection. Which TWO fix it?
- A. Tool search with defer_loading: true to load only relevant tools.
- B. Package reference material as Skills loaded on demand.
- C. Enlarge the window to 1M and keep loading everything.
- D. Escalate every request to Opus 5.
- E. Disable prompt caching.
Show answer
Answer: A and B.
Progressive discovery keeps context lean, restores the cache prefix, and improves tool selection. A bigger window (C) still pays for bloat; escalating (D) raises cost; disabling caching (E) is the opposite of the fix.
Q31D3 · Integration (incl. RAG)Select two
Which observability signals let you reconstruct a failed multi-step request end to end?
- A. A correlation ID threaded through app → model → tools → downstream.
- B. Per-span traces with token/cost telemetry, retrieved chunk IDs, and stop reasons.
- C. Only the final HTTP status code.
- D. Aggregate daily request counts only.
- E. The model's self-assessment of the request.
Show answer
Answer: A and B.
Correlation IDs plus per-span traces (with retrieval and cost detail) make an incident reconstructable. A status code (C) and daily counts (D) are too coarse; self-assessment (E) is unreliable.
Q32D4 · Evaluation, Testing and OptimizationSelect one
A bot reports 90% overall accuracy but only refund tickets fail. What is needed?
- A. Nothing; 90% is fine.
- B. Per-segment metrics that break accuracy down by ticket type, exposing the refund failure the aggregate hides.
- C. A bigger same-mix golden set.
- D. A higher-effort model for all tickets.
Show answer
Answer: B.
Aggregate masks a segment failure; stratified metrics expose it. The aggregate (A) misleads; a same-mix set (C) still hides it; blanket escalation (D) overpays.
Q33D4 · Evaluation, Testing and OptimizationSelect one
Grading answers with the same model and session that produced them is unsound because…
- A. it costs too much.
- B. same-session self-review carries the producing bias; the judge must be a separate model/session and calibrated against humans.
- C. the judge must always be larger.
- D. LLM-as-judge is never valid.
Show answer
Answer: B.
Independence plus human calibration is required. Cost (A) isn't the core issue; the judge needn't be larger (C); LLM-as-judge is valid when independent (D).
Q34D4 · Evaluation, Testing and OptimizationSelect one
Prompt B beats A on 7 of 12 hand-picked examples. Correct conclusion?
- A. Ship B.
- B. The sample is too small; run a paired A/B on a large per-segment set and test for statistical significance before promoting.
- C. Ship A as incumbent.
- D. Average the prompts.
Show answer
Answer: B.
7/12 is noise; promotion needs a significant paired result with no per-segment regression. Eyeballing (A, C) and averaging prompts (D) are unsound.
Q35D4 · Evaluation, Testing and OptimizationSelect one
Mean latency is 2 s but users complain; p95 is 12 s. What should the SLA track?
- A. Mean only.
- B. Tail latency (p95/p99), which reflects the slow experiences the mean hides.
- C. Request count.
- D. Token count.
Show answer
Answer: B.
Interactive SLAs track p95/p99. The mean (A) hides the tail; count (C) and tokens (D) aren't latency SLAs.
Q36D4 · Evaluation, Testing and OptimizationSelect two
An LLM judge rates longer answers higher regardless of correctness. What is it and the fix?
- A. Length bias.
- B. Calibrate against human labels and revise the rubric to score correctness/grounding, not length.
- C. The judge is reliable.
- D. Always prefer the longer answer.
- E. Delete the eval.
Show answer
Answer: A and B.
Length bias is fixed by calibration and a correctness-focused rubric. The judge isn't reliable (C), rewarding length (D) is the bug, and deleting evals (E) abandons measurement.
Q37D4 · Evaluation, Testing and OptimizationSelect one
How to prevent a fix for one segment from silently breaking another?
- A. Manual spot-checks post-release.
- B. A per-segment regression suite in CI that blocks merge on any segment regression.
- C. Trust the author.
- D. Test only the changed segment.
Show answer
Answer: B.
A per-segment regression gate in CI is the guard. Spot-checks (A) and author trust (C) are unreliable; testing only the changed segment (D) is the risk.
Q38D4 · Evaluation, Testing and OptimizationSelect one
A system fails only on the hardest 5% of cases. Most likely cause and fix?
- A. Retrieval; re-chunk everything.
- B. Model mismatch (tier too small); escalate just those cases to a stronger model via a cascade.
- C. Prompt failure; rewrite the whole prompt.
- D. Infra; add replicas.
Show answer
Answer: B.
'Only the hardest cases fail' signals model capability; a cascade escalates those cases. Re-chunking (A) and rewriting the prompt (C) target working layers; replicas (D) don't affect correctness.
Q39D4 · Evaluation, Testing and OptimizationSelect one
Cut cost 50% on a nightly bulk job with no latency need. Best lever and what must follow?
- A. Lower effort blindly and ship.
- B. Move to the Message Batches API (50% off, within 24 h) and re-run the per-segment regression suite to confirm no quality regression.
- C. Switch interactive traffic to Haiku too.
- D. Disable evaluation.
Show answer
Answer: B.
Latency-tolerant bulk work fits Batch; every cost change is followed by per-segment re-validation. Blind effort cuts (A) risk quality; changing interactive traffic (C) is out of scope; disabling evals (D) removes the safety net.
Q40D4 · Evaluation, Testing and OptimizationSelect two
Which two are the right metrics for a high-volume, cost-sensitive service with an interactive SLA?
- A. Cost per task.
- B. p95 latency.
- C. Mean latency only.
- D. Total fleet tokens.
- E. Prompt-version count.
Show answer
Answer: A and B.
Cost-sensitive + interactive → cost per task and p95. Mean (C) hides the tail; total tokens (D) and version count (E) aren't user-facing SLA metrics.
Q41D4 · Evaluation, Testing and OptimizationSelect one
A golden set is 85% English tickets; production fails on Spanish. Flaw and fix?
- A. The set was too small.
- B. It lacked per-segment coverage; stratify by language and report per segment.
- C. The model is broken.
- D. Online evals are unnecessary.
Show answer
Answer: B.
An unstratified set makes a segment invisible offline. Size (A) isn't the issue; the model isn't broken (C); online evals (D) are still needed but the root cause is coverage.
Q42D5 · Governance, Safety and Risk ManagementSelect one
A payment agent must never move over $10,000 without approval. Correct design?
- A. A firm system-prompt sentence.
- B. A tool-permission hook / validation that blocks transfers over \$10,000 and routes them to a human approval gate.
- C. Ask the model to report confidence before transfers.
- D. Escalate only when the user sounds anxious.
Show answer
Answer: B.
Hard limits need deterministic enforcement plus a human gate. Prompt wording (A), self-report (C), and sentiment (D) are anti-patterns.
Q43D5 · Governance, Safety and Risk ManagementSelect one
A US federal agency requires FedRAMP High. Compliant deployment?
- A. Direct Anthropic API.
- B. Claude via Bedrock or Vertex within the FedRAMP High boundary.
- C. Any cloud; Claude is inherently compliant.
- D. A personal laptop.
Show answer
Answer: B.
FedRAMP High is provided through Bedrock/Vertex boundaries. Direct API (A) is outside it; compliance isn't inherent (C); a laptop (D) is unacceptable.
Q44D5 · Governance, Safety and Risk ManagementSelect two
An agent summarising web pages reads hidden text telling it to email data externally. What is it and the defence?
- A. Indirect prompt injection.
- B. Treat external content as untrusted, apply least privilege so it lacks an unrestricted email tool, and validate/deny such actions.
- C. Trust it because it came via a tool.
- D. Raise effort.
- E. Add the instruction to the system prompt.
Show answer
Answer: A and B.
Malicious instructions in fetched content are indirect injection; defence is untrusted-content handling plus least privilege and validation. Trusting tool output (C) is the vulnerability; effort (D) is irrelevant; adding it to the prompt (E) executes the attack.
Q45D5 · Governance, Safety and Risk ManagementSelect two
A clinic wants a Claude assistant reading patient records. First compliance requirements?
- A. A signed BAA with the provider.
- B. PHI-handling controls (minimum necessary, access controls, audit logs) and an appropriate retention/ZDR posture.
- C. Nothing; it is internal.
- D. Publishing patient data to improve the model.
- E. Escalating on patient sentiment.
Show answer
Answer: A and B.
PHI under HIPAA requires a BAA and PHI controls with auditing. 'Internal' (C) doesn't exempt PHI; publishing data (D) is a violation; sentiment escalation (E) is unrelated.
Q46D5 · Governance, Safety and Risk ManagementSelect one
Escalating to a human whenever the customer 'sounds angry' is flawed because…
- A. it is optimal.
- B. sentiment is not a proxy for complexity/risk; escalate on measured complexity or deterministic rules instead.
- C. it escalates too rarely.
- D. anger always means the model failed.
Show answer
Answer: B.
Sentiment-based escalation conflates emotion with difficulty. It isn't optimal (A); frequency (C) isn't the core flaw; anger doesn't imply model failure (D).
Q47D5 · Governance, Safety and Risk ManagementSelect one
Which set best represents layered guardrails?
- A. One very detailed system prompt.
- B. Input classifier → prompt guidance → tool-permission hooks → output validation → human review for high-stakes actions.
- C. Only human review at the end.
- D. Only an output regex.
Show answer
Answer: B.
Defence in depth stacks independent layers. A single prompt (A), review-only (C), or regex-only (D) are single points of failure.
Q48D5 · Governance, Safety and Risk ManagementSelect two
High-risk processing of EU residents' personal data — required before launch?
- A. A DPIA.
- B. EU residency and a data-subject erasure/deletion path.
- C. Wait for a complaint.
- D. Store all data indefinitely.
- E. Escalate on model confidence.
Show answer
Answer: A and B.
GDPR high-risk processing requires a DPIA plus residency and DSAR/erasure support. Waiting (C) and indefinite storage (D) violate GDPR; confidence escalation (E) is unrelated.
Q49D5 · Governance, Safety and Risk ManagementSelect one
A prompt-injection incident is detected in production. Sound first containment?
- A. Delete all logs.
- B. Roll back the prompt/model version and/or disable the exploited tool via its permission hook, scoping impact with correlation IDs and traces.
- C. Ignore it until the next release.
- D. Publicly post customer data for transparency.
Show answer
Answer: B.
Containment is fast rollback and disabling the exploited capability, scoped via traces. Deleting logs (A) destroys evidence; ignoring (C) prolongs harm; disclosing data (D) is a second breach.
Q50D5 · Governance, Safety and Risk ManagementSelect one
Where should API keys and DB credentials for a Claude agent live?
- A. In the system prompt.
- B. In environment variables or a secret manager, never in prompts, CLAUDE.md, or logs.
- C. In CLAUDE.md for convenience.
- D. Hard-coded in tool source and printed to logs.
Show answer
Answer: B.
Secrets belong in env/secret managers, never in model-visible config or logs. The others create exfiltration paths.
Q51D5 · Governance, Safety and Risk ManagementSelect two
A hiring-screening assistant makes consequential recommendations. Required controls?
- A. Per-group fairness evaluation and symmetric treatment reviewed for bias.
- B. A human decision-maker retains authority over the hiring decision.
- C. Full automation to remove human bias.
- D. Escalate only when a candidate sounds confident.
- E. Store all applicant data indefinitely.
Show answer
Answer: A and B.
Consequential people-decisions require fairness evaluation and human authority. Full automation (C) launders bias; sentiment escalation (D) and indefinite retention (E) are anti-patterns/GDPR issues.
Q52D6 · Stakeholder Communication and Lifecycle ManagementSelect one
Two stakeholders disagree and no success metric exists; the team wants to build. First action?
- A. Build the most likely interpretation.
- B. Run structured discovery to align on measurable per-segment criteria and constraints, and get sign-off before design.
- C. Pick Opus 5 and start.
- D. Escalate to the vendor.
Show answer
Answer: B.
Without agreement or metrics the discovery gate isn't passed. Building (A), picking a model (C), or escalating (D) skip anchoring.
Q53D6 · Stakeholder Communication and Lifecycle ManagementSelect one
Presenting a workflow-vs-agentic decision to an executive sponsor AND engineers. Best approach?
- A. Send both the full technical ADR.
- B. Sponsor gets a one-page decision matrix with cost/risk/value; engineers get the ADR with rejected alternatives and consequences.
- C. Both get a value-only slide.
- D. Both get raw eval logs.
Show answer
Answer: B.
Communication matches audience altitude. One document for both (A, C) or raw logs (D) mismatches an audience.
Q54D6 · Stakeholder Communication and Lifecycle ManagementSelect one
Turning 'faster support' into an SLA. Correct approach?
- A. Promise 'much faster'.
- B. Agree numeric per-segment targets (p95 latency, deflection, cost/ticket) and report on a cadence.
- C. Track mean latency only.
- D. Leave it undefined.
Show answer
Answer: B.
SLAs must be measurable and per-segment with reporting. Vague promises (A) can't be verified; mean-only (C) hides the tail; undefined (D) invites disputes.
Q55D6 · Stakeholder Communication and Lifecycle ManagementSelect one
Which C4 level best shows executives how the system fits its environment?
- A. Level 4 (Code).
- B. Level 1 (Context).
- C. Level 3 (Component).
- D. Exhaustive sequence diagrams.
Show answer
Answer: B.
The Context level shows the system, users, and external systems for a broad audience. Code (A) and Component (C) are for engineers; exhaustive diagrams (D) overwhelm.
Q56D6 · Stakeholder Communication and Lifecycle ManagementSelect two
A model the service pins was retired. Correct migration approach?
- A. Read the deprecation notice, identify breaking changes, and re-run the regression suite per segment on the target model, adjusting prompts.
- B. Canary the new model with rollback, then ramp, and communicate timeline/differences.
- C. Swap the ID in production and hope.
- D. Keep calling the retired model.
- E. Disable evals to migrate faster.
Show answer
Answer: A and B.
Deprecation is a governed lifecycle event: assess, re-validate per segment, canary, communicate. A blind swap (C) risks regressions; calling a retired model (D) fails; disabling evals (E) removes the safety net.
Q57D6 · Stakeholder Communication and Lifecycle ManagementSelect one
What deliverable is required to pass the handoff gate to operations?
- A. A marketing deck.
- B. A runbook (dashboards, alerts, on-call, rollback, dependencies) plus C4 docs so operators can run and recover it.
- C. Nothing; they'll figure it out.
- D. Only the source code.
Show answer
Answer: B.
Handoff's exit criterion is that operators can run and recover the system. A deck (A), nothing (C), or code alone (D) don't enable safe operation.
Q58D6 · Stakeholder Communication and Lifecycle ManagementSelect one
When should a DPIA/BAA be addressed in the lifecycle?
- A. After launch if a regulator asks.
- B. At the design gate, so residency, retention, and PHI/PII handling shape the architecture before build.
- C. Never, if internal.
- D. Only during incident response.
Show answer
Answer: B.
Compliance must shape design, so DPIA/BAA belong to the design exit criterion. Later (A, D) is too late; 'internal' (C) doesn't exempt regulated data.
Q59D6 · Stakeholder Communication and Lifecycle ManagementSelect two
A technically strong system sees low adoption. Which levers help most?
- A. Training/enablement and clear communication of capabilities and limits.
- B. Phased rollout with feedback capture and success-metric reporting.
- C. Mandatory immediate use with no support.
- D. Removing human-in-the-loop gates for speed.
- E. Hiding limitations from users.
Show answer
Answer: A and B.
Adoption is driven by enablement, phased rollout, feedback, and demonstrated value. Forced use (C), removing gates (D), and hiding limits (E) backfire.
Q60D6 · Stakeholder Communication and Lifecycle ManagementSelect two
The team plans to deploy to 100% of users with no runbook and no agreed metrics. What does the correct answer insert?
- A. Agreed measurable success criteria before proceeding.
- B. A runbook and a canary/phased rollout with rollback before full deployment.
- C. Immediate full deployment to gather data.
- D. Skipping monitoring to reduce overhead.
- E. Per-engineer metric definitions.
Show answer
Answer: A and B.
The missing gates are agreed criteria and a runbook plus canary with rollback. Full deployment (C) is high blast radius; skipping monitoring (D) blinds ops; per-engineer metrics (E) destroy a shared definition of success.
Q61D7 · Developer Productivity and Operational EnablementSelect one
A platform team wants a coding standard and denied destructive commands enforced across all repos with no override. Best mechanism?
- A. Ask developers to add rules to CLAUDE.local.md.
- B. Managed/enterprise policy plus checked-in ./CLAUDE.md and .claude/settings.json with permissions.deny.
- C. A shared Slack message.
- D. A quarterly email.
Show answer
Answer: B.
Non-overridable, org-wide enforcement is what managed policy provides, with checked-in project config as the baseline. CLAUDE.local.md (A) is git-ignored/overridable; Slack (C) and email (D) aren't enforcement.
Q62D7 · Developer Productivity and Operational EnablementSelect one
Setting up automated diff review in CI. Correct approach?
- A. Run interactive Claude and paste diffs by hand.
- B. Headless mode: claude -p '…' --output-format json with restricted --allowedTools and an appropriate --permission-mode.
- C. Give the CI agent all tools.
- D. Disable permissions in CI.
Show answer
Answer: B.
CI is non-interactive → headless with structured output and restricted tools. Interactive use (A) can't run in CI; all-tools (C) and disabled permissions (D) break least privilege.
Q63D7 · Developer Productivity and Operational EnablementSelect one
A manager wants to measure AI productivity by lines of code generated and suggestions accepted. Guidance?
- A. Those are good primary metrics.
- B. Measure delivery outcomes (cycle time, change-failure/rework rate, review quality); lines and acceptances are vanity metrics.
- C. Measure tokens consumed.
- D. Don't measure at all.
Show answer
Answer: B.
Productivity ties to delivery outcomes, not volume. Lines/acceptances (A) and tokens (C) are vanity signals; not measuring (D) forfeits value demonstration.

Instructions

Domain distribution

Score interpretation

Take the practice exam

Take the practice exam

By domain

Correction

All questions (review mode)