{ "id": "claude-ccar-p-mock-01", "version": 2, "track": "claude-ccar-p", "kind": "mock", "title": "Architect Professional Full Mock", "timeLimitMinutes": 120, "questions": [ { "id": "ccar-p-m1-001", "domain": "solution-design-architecture", "objective": "Translate business problems into Claude solutions", "type": "single", "prompt": "A claims director asks for a fully autonomous system that will reduce handling time, but supplies no baseline, case boundaries, authority limits, or acceptable error rate. What should the architect do first?", "options": [ "Promise a latency target based on a vendor benchmark", "Select the highest-capability Claude model and build a read-only pilot on recent claims, using reviewer agreement to infer suitable case and authority boundaries", "Design a manager agent with one subagent per claim type", "Run structured discovery to define the current workflow, measurable outcome, risk boundaries, authority, and non-goals" ], "correct": [ 3 ], "explanation": "Architecture cannot be evaluated until the business outcome, workflow, error cost, and action authority are explicit. Structured discovery turns a feature request into testable requirements and exposes whether autonomy is justified.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://www.anthropic.com/research/building-effective-agents", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-002", "domain": "solution-design-architecture", "objective": "Align designs to value, cost, performance, and SLAs", "type": "multiple", "prompt": "A support drafting system is being designed for a premium service. Which requirements belong in the initial architecture decision? Select all that apply.", "options": [ "Task-quality target on a representative ticket population", "Require one model across all environments for consistency", "Maximum action authority and human-approval triggers", "P95 response latency and expected request volume" ], "correct": [ 0, 2, 3 ], "explanation": "Quality, performance, scale, and action authority directly constrain the design. A model tier is an implementation choice to test against those requirements, not a business requirement by itself.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://platform.claude.com/docs/en/about-claude/models/overview", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-003", "domain": "solution-design-architecture", "objective": "Choose workflow, agentic, or augmented-LLM patterns", "type": "single", "prompt": "A loan-document process always extracts fields, validates them against deterministic rules, checks two databases, and routes exceptions to a reviewer. Which pattern is strongest?", "options": [ "A deterministic workflow with bounded Claude extraction and reasoning steps", "An unrestricted agent with access to every banking system", "One structured prompt that includes the documents and copied database values, asks Claude to execute every rule, and routes low-confidence outputs to review", "A peer-to-peer swarm with no coordinator" ], "correct": [ 0 ], "explanation": "The sequence and controls are known, so deterministic orchestration provides clearer auditability and failure handling. Agentic planning should be reserved for paths that genuinely depend on evidence discovered during execution.", "references": [ "https://www.anthropic.com/research/building-effective-agents", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-004", "domain": "solution-design-architecture", "objective": "Design input, processing, output, and feedback loops", "type": "single", "prompt": "A contract-review application produces good summaries but does not improve after reviewers correct missed clauses. Which architectural component is missing?", "options": [ "A larger context window", "A governed feedback loop that captures corrections, labels failure classes, and feeds evaluation and iteration", "A second output formatter", "A reviewer dashboard that displays missed-clause feedback beside each summary but does not convert corrections into governed labels or evaluation cases" ], "correct": [ 1 ], "explanation": "Reviewer corrections create value only when they enter a controlled learning and evaluation loop. The system needs provenance, labels, ownership, and release criteria rather than an unstructured store of comments.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-005", "domain": "solution-design-architecture", "objective": "Design multi-agent orchestration", "type": "multiple", "prompt": "A due-diligence system uses a manager and four specialist agents. Which controls make the hierarchy defensible? Select all that apply.", "options": [ "Let specialists edit one shared report while the manager infers ownership", "Give each specialist a bounded objective, source set, and output contract", "Propagate correlation IDs and budget limits through every delegation", "Use a separate evaluator to verify evidence before the manager synthesizes" ], "correct": [ 1, 2, 3 ], "explanation": "Bounded contracts, independent evidence checks, traceability, and inherited budgets keep delegation inspectable. Silent cross-editing destroys ownership and makes failures difficult to attribute or reproduce.", "references": [ "https://www.anthropic.com/research/building-effective-agents", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-006", "domain": "solution-design-architecture", "objective": "Decompose complex problems", "type": "single", "prompt": "A migration plan has discovery, schema design, transformation, verification, and cutover stages. Which decomposition is safest?", "options": [ "Start cutover first so the team learns faster", "Give one planning agent the full migration brief and authority to reorder stages whenever new evidence appears, while relying on its session memory for predecessor status", "Model the dependencies, parallelize only independent discovery tasks, and gate transformation and cutover on verified predecessors", "Run every stage in parallel to minimize elapsed time" ], "correct": [ 2 ], "explanation": "Decomposition must preserve causal dependencies. Independent work may proceed concurrently, while destructive or irreversible stages require accepted inputs, checkpoints, and explicit go or no-go gates.", "references": [ "https://www.anthropic.com/research/building-effective-agents", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-007", "domain": "solution-design-architecture", "objective": "Align designs to value, cost, performance, and SLAs", "type": "single", "prompt": "Two architectures have the same answer accuracy. Design A costs half as much but sends twice as many cases to human review, making total handling cost higher. Which metric should drive the decision?", "options": [ "Output token count alone", "Model inference cost alone", "Total automated handling time before human review, because the architecture with lower model cost also produces fast preliminary decisions", "Total cost per successfully completed business outcome, including review and rework" ], "correct": [ 3 ], "explanation": "Architecture economics must include the entire workflow. Cheap inference can be a false economy when it increases review labor, escalations, retries, or downstream correction.", "references": [ "https://platform.claude.com/docs/en/about-claude/pricing", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-008", "domain": "solution-design-architecture", "objective": "Design input, processing, output, and feedback loops", "type": "multiple", "prompt": "A customer-facing Claude service has a strict availability objective. Which degradation paths should be designed before launch? Select all that apply.", "options": [ "Queues and bounded retries for transient provider failures", "Fallback to a narrower read-only workflow when optional tools fail", "Explicit unavailable or escalate states instead of fabricated answers", "Retry failed requests indefinitely so availability remains high" ], "correct": [ 0, 1, 2 ], "explanation": "Graceful degradation preserves safe value under partial failure. Explicit states, queues, and bounded retries support the service objective; unlimited retries create load amplification and uncontrolled latency.", "references": [ "https://sre.google/sre-book/table-of-contents/", "https://platform.claude.com/docs/en/api/messages", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-009", "domain": "solution-design-architecture", "objective": "Design input, processing, output, and feedback loops", "type": "single", "prompt": "An application validates model output but never checks whether a requested refund actually appears in the payment system. Which layer is missing?", "options": [ "Independent final-state verification against the system of record", "A richer refund-tool response containing the requested amount, model confidence, policy citations, and a success status returned by the orchestration layer", "Prompt caching", "A prompt template" ], "correct": [ 0 ], "explanation": "Output validation proves only the response contract. Agentic applications also require a final-state check that confirms the external action and catches confident but unexecuted claims.", "references": [ "https://www.anthropic.com/research/building-effective-agents", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-010", "domain": "solution-design-architecture", "objective": "Choose workflow, agentic, or augmented-LLM patterns", "type": "single", "prompt": "A global service requires regional data residency and can tolerate a 30-minute recovery point for its knowledge index. Which design is most defensible?", "options": [ "Keep the regulated source in approved regional storage but query it through one centrally managed cross-region index to simplify recovery and relevance tuning", "Map data classes and provider regions, keep regulated data in approved boundaries, and design regional indexes with tested recovery", "Copy every source into one global prompt", "Ignore residency because prompts are transient" ], "correct": [ 1 ], "explanation": "Residency and recovery requirements shape storage, retrieval, provider selection, and failover. The design must document approved regions and prove restoration instead of assuming transient model calls remove data obligations.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://trust.anthropic.com/", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-011", "domain": "solution-design-architecture", "objective": "Translate business problems into Claude solutions", "type": "single", "prompt": "An architecture review compares workflow and agent designs but records only the chosen option. What should be added?", "options": [ "The architect's private chain of thought", "A target-state diagram and implementation plan that describe the selected option in enough detail to recreate it if the current architecture owner leaves", "A durable decision record containing context, constraints, rejected alternatives, consequences, verification, and reversal conditions", "A screenshot of the whiteboard only" ], "correct": [ 2 ], "explanation": "A professional decision record preserves why the choice fit current evidence and what would trigger reconsideration. This supports implementation, review, operations, and later reversal without exposing private reasoning.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-012", "domain": "models-prompting-context", "objective": "Select models from explicit tradeoffs", "type": "single", "prompt": "A smaller Claude model and a larger model both pass safety gates, but the larger model improves complex reasoning accuracy by 6 points at three times the cost. What should the architect do?", "options": [ "Use the smaller model for every request because it passes safety gates, then add a human review rule for the complex cases instead of routing by measured value", "Average the two model outputs without validation", "Use the larger model for every request", "Segment the workload, quantify business value of the gain, and route only cases that earn the larger model" ], "correct": [ 3 ], "explanation": "Model selection is workload-specific. Routing preserves economical handling for simple cases while applying additional capability where the measured quality improvement justifies cost and latency.", "references": [ "https://platform.claude.com/docs/en/about-claude/models/overview", "https://platform.claude.com/docs/en/about-claude/pricing", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-013", "domain": "models-prompting-context", "objective": "Select models from explicit tradeoffs", "type": "multiple", "prompt": "A model-routing architecture is proposed. Which controls are required? Select all that apply.", "options": [ "A representative eval for the router's classification decisions", "Cost and latency telemetry by route", "Route uncertain cases to the cheapest model before escalating", "Fallback behavior when the selected model is unavailable or fails its contract" ], "correct": [ 0, 1, 3 ], "explanation": "The router becomes a production decision component and needs its own quality, fallback, and economics evidence. Uncertain high-impact cases generally need stronger review or capability, not automatic cost minimization.", "references": [ "https://platform.claude.com/docs/en/about-claude/models/overview", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-014", "domain": "models-prompting-context", "objective": "Design prompts, templates, and guardrails", "type": "single", "prompt": "A production prompt has grown to 9,000 tokens of duplicated policy, examples, and task-specific rules. What is the strongest redesign?", "options": [ "Separate always-needed policy, reusable task modules, scoped examples, and deterministic controls, then evaluate each change", "Remove all constraints and rely on a model upgrade", "Keep one complete prompt but deduplicate exact repeated lines, group the remaining policy and examples by topic, and rely on prompt caching to control cost", "Move every line into the user message" ], "correct": [ 0 ], "explanation": "Modular instruction architecture reduces context cost and conflict. The team should preserve core policy, load specialized procedure only when relevant, enforce invariants outside the prompt, and validate behavior through evals.", "references": [ "https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview", "https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-015", "domain": "models-prompting-context", "objective": "Apply zero-shot, few-shot, and reasoning techniques", "type": "single", "prompt": "A claims classifier performs well on obvious cases but confuses borderline exclusions. Which prompting intervention is most targeted?", "options": [ "Increase output length to 5,000 words", "Add few-shot examples drawn from the difficult boundary cases with correct rationale and labels", "Add a larger balanced sample of routine approved and denied claims so the model sees the overall label distribution, without emphasizing borderline exclusions", "Tell the model it is a world-class classifier" ], "correct": [ 1 ], "explanation": "Few-shot examples are most valuable where category boundaries are ambiguous. Representative hard cases teach the intended distinction, while easy repetition and personas do not address the observed failure.", "references": [ "https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-016", "domain": "models-prompting-context", "objective": "Manage context windows and token use", "type": "single", "prompt": "A research agent repeatedly includes 200 pages of source material even though each question needs only a few sections. What is the best architecture?", "options": [ "Pre-summarize each source into a short paragraph, concatenate all summaries into a stable context, and retrieve from that compressed corpus without source-level provenance", "Always select the largest context window", "Retrieve and rank relevant sections, preserve provenance, and keep durable state outside the prompt", "Drop source citations to save tokens" ], "correct": [ 2 ], "explanation": "Context should carry the evidence needed for the current decision, not the entire corpus by default. Retrieval and durable state reduce bloat while provenance preserves verification.", "references": [ "https://platform.claude.com/docs/en/build-with-claude/context-windows", "https://www.anthropic.com/news/contextual-retrieval", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-017", "domain": "models-prompting-context", "objective": "Reuse stable prompts with caching, modular prompts, and Skills", "type": "multiple", "prompt": "A service repeats a long stable policy and tool catalog across requests. Which choices improve cache economics? Select all that apply.", "options": [ "Place the stable prefix before customer-specific content", "Put tenant metadata first to bind every cached prefix", "Measure cache creation, hit rate, latency, and cost", "Keep reusable prefix bytes stable where practical" ], "correct": [ 0, 2, 3 ], "explanation": "Prompt caching benefits from stable shared prefixes and empirical measurement. Volatile content near the beginning prevents reuse for everything that follows it.", "references": [ "https://platform.claude.com/docs/en/build-with-claude/prompt-caching", "https://platform.claude.com/docs/en/about-claude/pricing", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-018", "domain": "models-prompting-context", "objective": "Reuse stable prompts with caching, modular prompts, and Skills", "type": "multiple", "prompt": "Which content is best packaged as an Agent Skill instead of always placed in the root prompt? Select all that apply.", "options": [ "A universal safety rule that loads only after model risk detection", "Reference files and scripts used by one specialized workflow", "A domain playbook that should load when its task is detected", "A reusable incident-analysis procedure needed only during incidents" ], "correct": [ 1, 2, 3 ], "explanation": "Skills support progressive disclosure for reusable specialized procedure and assets. A safety invariant needed every time belongs in trusted core policy and deterministic controls.", "references": [ "https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-019", "domain": "models-prompting-context", "objective": "Apply zero-shot, few-shot, and reasoning techniques", "type": "single", "prompt": "Extended thinking improves one architecture-planning eval but causes the interactive service to miss its latency objective. What is the strongest response?", "options": [ "Keep extended thinking on the synchronous route and stream a generated progress explanation to users, because perceived responsiveness may offset the measured latency miss", "Disable all reasoning features permanently", "Ignore the latency objective", "Use the measured gain only in an asynchronous or high-complexity route whose service objective can absorb it" ], "correct": [ 3 ], "explanation": "Reasoning budget must fit the service and task. Routing difficult or asynchronous work to a higher-effort configuration preserves quality without violating the primary interactive path.", "references": [ "https://platform.claude.com/docs/en/build-with-claude/extended-thinking", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/01-claude-product-and-model-landscape" ] }, { "id": "ccar-p-m1-020", "domain": "integration", "objective": "Reduce capability bloat", "type": "multiple", "prompt": "An enterprise agent loads 140 overlapping tools on every turn. Which redesigns are appropriate? Select all that apply.", "options": [ "Group capabilities into coherent domain surfaces", "Use progressive discovery so detailed tools load only when relevant", "Remove or merge overlapping tools based on trajectory evals", "Keep all tools but shorten each description to one sentence" ], "correct": [ 0, 1, 2 ], "explanation": "Capability bloat consumes context and increases selection ambiguity. Domain scoping, progressive discovery, and evidence-driven consolidation preserve useful reach with a smaller active surface.", "references": [ "https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview", "https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-021", "domain": "integration", "objective": "Find authentication and authorization gaps", "type": "multiple", "prompt": "An MCP tool accepts tenant_id and user_role directly from model arguments. Which corrections are required? Select all that apply.", "options": [ "Authorize the specific operation and resource on the server", "Bind tenant and identity from the authenticated session", "Trust schema-validated tenant and role arguments from the model", "Treat model-provided role as advisory only, never proof of authority" ], "correct": [ 0, 1, 3 ], "explanation": "Schema validity does not establish identity or permission. Authentication comes from trusted session state, and server-side authorization must constrain every resource and action.", "references": [ "https://modelcontextprotocol.io/specification/latest/basic/authorization", "https://modelcontextprotocol.io/docs/tutorials/security/security_best_practices", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-022", "domain": "integration", "objective": "Design RAG chunking and indexing", "type": "single", "prompt": "A corpus contains long legal documents whose meaning depends on sections, clauses, and cross-references. Which chunking strategy is strongest?", "options": [ "Structure-aware chunks that retain section metadata, overlap where needed, and preserve document and clause identifiers", "Fixed five-character chunks", "Sentence-level chunks grouped into fixed windows after shuffling boilerplate clauses away, with the parent document ID retained for reconstruction", "Uniform token-sized chunks with generous overlap and document IDs, because consistent embedding units simplify indexing and cross-document retrieval at scale" ], "correct": [ 0 ], "explanation": "Structure-aware chunking preserves semantic units and provenance while keeping retrieval granular. Metadata and controlled overlap support cross-reference resolution and later citation checks.", "references": [ "https://www.anthropic.com/news/contextual-retrieval", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-023", "domain": "integration", "objective": "Match retrieval to data shape and query pattern", "type": "single", "prompt": "A support assistant retrieves an obsolete cancellation policy even though the new policy is indexed. Which system change most directly addresses the defect?", "options": [ "Increase semantic-ranking weight for cancellation queries and ask the generator to compare any retrieved policy dates before choosing which passage to cite", "Add version, effective-date, and authority metadata; filter or rank to the active policy and test freshness", "Increase generation temperature", "Remove all document identifiers" ], "correct": [ 1 ], "explanation": "This is a retrieval-governance failure. Explicit authority and temporal metadata let the retriever select the active source and make freshness observable in tests and operations.", "references": [ "https://www.anthropic.com/news/contextual-retrieval", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-024", "domain": "integration", "objective": "Match retrieval to data shape and query pattern", "type": "multiple", "prompt": "A product search must handle exact part numbers and natural-language descriptions. Which retrieval design is appropriate? Select all that apply.", "options": [ "Lexical retrieval for exact identifiers", "Recall likely catalog entries before retrieval", "A reranking stage evaluated on representative queries", "Semantic retrieval for paraphrased needs" ], "correct": [ 0, 2, 3 ], "explanation": "Exact and semantic queries have different strengths, so hybrid retrieval plus reranking can serve both. Product availability and identifiers require current authoritative data, not model memory.", "references": [ "https://www.anthropic.com/news/contextual-retrieval", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-025", "domain": "integration", "objective": "Use progressive discovery instead of monolithic context where appropriate", "type": "multiple", "prompt": "A large internal platform has hundreds of APIs. Which progressive-discovery design is strongest? Select all that apply.", "options": [ "Put compressed summaries for every API in root context", "Record discovery and selection in the trace", "Expose a small capability index with concise domain descriptions", "Load detailed schemas only after a domain is selected" ], "correct": [ 1, 2, 3 ], "explanation": "Progressive discovery keeps initial context small while retaining access to deeper capability. Tracing makes routing errors diagnosable; monolithic schemas increase cost and confusion.", "references": [ "https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview", "https://modelcontextprotocol.io/specification", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-026", "domain": "integration", "objective": "Choose MCP, API, CLI, or agent-to-agent integration", "type": "single", "prompt": "One service makes one stable, high-volume call to an internal risk API. No other host needs discovery. Which integration is the best default?", "options": [ "A public MCP registry entry", "A small MCP server deployed beside the service so the internal risk capability can later be discovered by additional hosts without changing the client contract", "A direct typed API client with authentication, timeouts, and contract tests", "A new multi-agent protocol" ], "correct": [ 2 ], "explanation": "A direct typed client is the smallest reliable boundary for one stable high-volume integration. MCP or agent-to-agent protocols add value only when interoperability or delegated coordination justifies their lifecycle cost.", "references": [ "https://modelcontextprotocol.io/specification", "https://platform.claude.com/docs/en/api/messages", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-027", "domain": "integration", "objective": "Choose MCP, API, CLI, or agent-to-agent integration", "type": "single", "prompt": "A desktop host and two IDEs need the same governed repository tools and reusable resources. Which option best fits?", "options": [ "Duplicate bespoke connectors in each host", "A spreadsheet copied into every session", "Three host-specific connectors generated from one shared integration package, all calling the same governed repository API and consent service", "A reviewed MCP server with negotiated capabilities and host-specific consent" ], "correct": [ 3 ], "explanation": "MCP standardizes discovery and invocation across compliant hosts while preserving one governed server boundary. Authentication, authorization, consent, and versioning remain necessary.", "references": [ "https://modelcontextprotocol.io/specification", "https://modelcontextprotocol.io/docs/tutorials/security/security_best_practices", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-028", "domain": "integration", "objective": "Select observability at scale", "type": "multiple", "prompt": "A request crosses an API gateway, Claude call, MCP server, and database. Which telemetry is needed for diagnosis at scale? Select all that apply.", "options": [ "Latency and result class for each boundary", "Versioned model, prompt, tool, and policy identifiers", "A propagated correlation or trace ID", "Log bearer-token fingerprints in every trace span" ], "correct": [ 0, 1, 2 ], "explanation": "Correlation and versioned boundary telemetry reconstruct the request path and identify regressions. Credentials and unrestricted sensitive content should be redacted or omitted from observability.", "references": [ "https://opentelemetry.io/docs/specs/otel/", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-029", "domain": "integration", "objective": "Balance accuracy and latency", "type": "single", "prompt": "A RAG pipeline improves recall by issuing five query expansions, but p95 latency exceeds the user objective and answer quality improves only 0.4 points. What should the architect do?", "options": [ "Use eval evidence to reduce or condition expansion, then measure quality and latency by query slice", "Remove retrieval entirely", "Hide the latency metric from stakeholders", "Keep all five expansions for every query but run them concurrently and cache repeated candidates, optimizing transport latency before changing retrieval behavior" ], "correct": [ 0 ], "explanation": "The small average quality gain may not earn the latency cost. Conditional expansion for difficult query slices can preserve value while meeting the service objective.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-030", "domain": "integration", "objective": "Find authentication and authorization gaps", "type": "single", "prompt": "A third-party model platform is proposed for regulated regional workloads. What must be verified before approval?", "options": [ "Only whether it exposes a compatible chat method", "Identity, data processing and residency, model and feature availability, network path, logging, retention, and contractual controls", "Whether its SDK preserves the existing method names, streaming events, and error classes so the application can migrate without code changes", "Whether a demo completes once" ], "correct": [ 1 ], "explanation": "API compatibility does not establish compliance or operational suitability. The architect must verify the full data and identity boundary and document which controls remain the customer's responsibility.", "references": [ "https://trust.anthropic.com/", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-031", "domain": "integration", "objective": "Choose MCP, API, CLI, or agent-to-agent integration", "type": "multiple", "prompt": "A tool times out after it may have created a customer record. Which integration controls are required before retry? Select all that apply.", "options": [ "A stable idempotency key", "A bounded retry policy based on failure class", "Ask a model whether the timed-out write succeeded", "Authoritative reconciliation of the operation outcome" ], "correct": [ 0, 1, 3 ], "explanation": "Ambiguous side effects require transaction-level evidence. Idempotency, reconciliation, and classified retry policy prevent duplicates; generation settings cannot determine whether the first write completed.", "references": [ "https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview", "https://sre.google/sre-book/table-of-contents/", "certifications/claude/lessons/04-context-knowledge-memory-and-caching" ] }, { "id": "ccar-p-m1-032", "domain": "evaluation-testing-optimization", "objective": "Define accuracy, latency, cost, safety, and security metrics", "type": "multiple", "prompt": "Which measures form a balanced production scorecard for an account-support agent? Select all that apply.", "options": [ "Cost per successful resolution and human-review minutes", "Track response length and model confidence", "P50 and P95 latency plus availability", "Task success and unsupported-claim rate" ], "correct": [ 0, 2, 3 ], "explanation": "A balanced scorecard covers quality, reliability, economics, and operational burden. Stylistic counts matter only if tied to a validated user requirement.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-033", "domain": "evaluation-testing-optimization", "objective": "Build representative datasets and mixed-method test frameworks", "type": "multiple", "prompt": "A team builds an evaluation set from production traffic. Which practices improve validity? Select all that apply.", "options": [ "Tune prompts against held-out failures until those cases pass", "Represent common, boundary, high-impact, and adversarial slices", "Preserve provenance and label quality while minimizing sensitive data", "Separate development cases from held-out release cases" ], "correct": [ 1, 2, 3 ], "explanation": "Representative slices and trustworthy labels make the set useful, while a held-out partition limits overfitting. Repeated tuning on release cases converts them into development data.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "https://www.nist.gov/itl/ai-risk-management-framework", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-034", "domain": "evaluation-testing-optimization", "objective": "Build representative datasets and mixed-method test frameworks", "type": "single", "prompt": "An LLM judge gives a helpfulness score. How should the team establish whether the judge is reliable enough for a release gate?", "options": [ "Ask the judge to score its own accuracy", "Use the same model and prompt as the generator", "Calibrate it against independently labeled human examples and inspect false positives, false negatives, and slice disagreement", "Accept it after prompt-level consistency tests show stable JSON scores across repeated samples, using disagreement with human labels only to refine the judge prompt" ], "correct": [ 2 ], "explanation": "Judge validity is empirical. Human calibration and disagreement analysis reveal rubric ambiguity and shared blind spots that structured output or self-assessment cannot remove.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-035", "domain": "evaluation-testing-optimization", "objective": "Diagnose prompt, hallucination, retrieval, and model failures", "type": "multiple", "prompt": "A RAG answer cites the wrong policy. Which tests isolate the cause? Select all that apply.", "options": [ "Evaluate reranker ordering independently", "Measure whether the correct document appeared in the retrieved top-k set", "Give Claude the correct source directly and test answer faithfulness", "Change prompt, reranker, and model together before retesting" ], "correct": [ 0, 1, 2 ], "explanation": "Separate retrieval, ranking, and generation tests localize the failure. Changing every component together may improve one run but destroys causal evidence.", "references": [ "https://www.anthropic.com/news/contextual-retrieval", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-036", "domain": "evaluation-testing-optimization", "objective": "Run A/B tests and iterative improvements", "type": "single", "prompt": "Variant B raises offline answer quality by 3 points but increases live abandonment. What should determine rollout?", "options": [ "Offline score alone", "The offline score if it was measured on a representative dataset, while treating abandonment as a temporary interface-adaptation effect after release", "The architect's personal preference", "A predeclared decision rule combining guardrails and the primary live outcome, followed by slice analysis" ], "correct": [ 3 ], "explanation": "Offline quality is one signal. A production experiment should use predeclared business and safety metrics, ensure valid assignment, and inspect important slices before deciding.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-037", "domain": "evaluation-testing-optimization", "objective": "Build representative datasets and mixed-method test frameworks", "type": "single", "prompt": "One evaluation run passes all 50 cases. What can the architect conclude?", "options": [ "Only that one sampled run passed; higher-variance or high-risk cases need repeated measurement and confidence analysis", "The model is deterministic", "The system will always pass those cases", "The current prompt-model combination is ready for launch on the represented cases, provided deterministic checks pass, without estimating run-to-run uncertainty" ], "correct": [ 0 ], "explanation": "Probabilistic model behavior and changing integrations make one run insufficient evidence of stability. Repeated trials and production monitoring quantify variance and detect distribution shift.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "https://platform.claude.com/docs/en/about-claude/models/overview", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-038", "domain": "evaluation-testing-optimization", "objective": "Diagnose prompt, hallucination, retrieval, and model failures", "type": "single", "prompt": "The provider wire response contains a citation field, but the application trace and UI do not. What should the team inspect first?", "options": [ "The UI rendering component and browser compatibility layer, because the trace may intentionally omit provider-only citation metadata before display", "SDK deserialization and each application serialization boundary", "The retrieval and citation-generation prompts, because a citation visible on the provider wire may still have been produced from a malformed evidence chunk", "The system prompt" ], "correct": [ 1 ], "explanation": "The field existed at the provider boundary, so the likely defect is integration mapping. Comparing wire, SDK object, service DTO, and UI payload locates where the field disappeared.", "references": [ "https://platform.claude.com/docs/en/api/messages", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-039", "domain": "evaluation-testing-optimization", "objective": "Optimize tokens, latency, and cost-performance", "type": "single", "prompt": "Prompt caching cuts input cost by 35 percent, but a stale policy prefix causes incorrect decisions. What is the correct optimization response?", "options": [ "Increase cache lifetime without measurement", "Remove correctness from the scorecard", "Restore source freshness and versioned invalidation, then evaluate cache economics on correct outputs", "Keep caching enabled but shorten the cache lifetime and add a prompt instruction to prefer newer retrieved policy when it conflicts with the stable prefix" ], "correct": [ 2 ], "explanation": "Optimization is subordinate to correctness. Stable prefixes need explicit content ownership and invalidation so cost savings do not preserve obsolete policy.", "references": [ "https://platform.claude.com/docs/en/build-with-claude/prompt-caching", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-040", "domain": "evaluation-testing-optimization", "objective": "Run A/B tests and iterative improvements", "type": "multiple", "prompt": "Which release gates protect against a misleading average improvement? Select all that apply.", "options": [ "No regression beyond threshold in high-impact slices", "Zero new unauthorized side effects", "Gate only on a significant aggregate improvement", "Cost and latency within approved budgets" ], "correct": [ 0, 1, 3 ], "explanation": "Non-negotiable safety failures, critical slices, and operational budgets prevent aggregate gains from hiding unacceptable regressions. Averages cannot express every risk boundary.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "https://www.nist.gov/itl/ai-risk-management-framework", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-041", "domain": "evaluation-testing-optimization", "objective": "Monitor with logs, traces, and observability", "type": "multiple", "prompt": "Which information belongs in a production trace for a tool-using Claude workflow? Select all that apply.", "options": [ "Model, prompt, schema, tool, and policy versions", "Encrypted full prompts, retrieved documents, and tool arguments in a restricted trace store so investigators can reproduce any result without source lookups", "Stop reason and redacted tool-decision events", "Correlation IDs and final-state verification result" ], "correct": [ 0, 2, 3 ], "explanation": "Versioned and correlated decision events support reproduction and incident analysis. Sensitive payloads and credentials should be minimized, redacted, and protected rather than copied into general telemetry.", "references": [ "https://opentelemetry.io/docs/specs/otel/", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/02-model-selection-and-token-economics" ] }, { "id": "ccar-p-m1-042", "domain": "governance-safety-risk", "objective": "Layer guardrails and safety controls", "type": "multiple", "prompt": "A coding agent can read files, execute commands, and access a package registry. Which layered controls are appropriate? Select all that apply.", "options": [ "A system instruction that assigns Claude the role of a security-conscious senior engineer and requires explicit reasoning before any tool with side effects", "Prompt-level trust-boundary instructions", "Pre-tool policy hooks and explicit approval for consequential actions", "Filesystem and network sandboxing with least privilege" ], "correct": [ 1, 2, 3 ], "explanation": "Defense in depth combines behavioral guidance, deterministic gates, human authority, and infrastructure isolation. Capability claims do not constrain execution or replace enforceable boundaries.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/mitigate-jailbreaks", "https://code.claude.com/docs/en/hooks-guide", "https://code.claude.com/docs/en/security", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-043", "domain": "governance-safety-risk", "objective": "Design human-in-the-loop validation", "type": "multiple", "prompt": "Which properties make a human-review step meaningful? Select all that apply.", "options": [ "A qualified reviewer with evidence and authority", "A defined trigger based on impact, uncertainty, or policy", "A time objective and safe fallback when review is unavailable", "A post-action reviewer checkpoint that can reverse most outcomes and records whether the original model action should remain in effect" ], "correct": [ 0, 1, 2 ], "explanation": "Human review needs a trigger, accountable role, evidence packet, time budget, and fallback. Review after an irreversible action cannot serve as preauthorization.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://www.nist.gov/itl/ai-risk-management-framework", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-044", "domain": "governance-safety-risk", "objective": "Account for GDPR, HIPAA, FedRAMP, and applicable regulation", "type": "multiple", "prompt": "A European customer-service assistant is being designed. Which GDPR-aligned architectural questions should be answered? Select all that apply.", "options": [ "What lawful basis and purpose govern processing", "How data-subject rights and deletion propagate through stores and logs", "Which optional customer fields could improve future personalization, provided the privacy notice describes broad service-improvement use and access is restricted", "Which personal data is necessary and how long it is retained" ], "correct": [ 0, 1, 3 ], "explanation": "Purpose limitation, data minimization, retention, and rights handling shape the data lifecycle. Speculative collection conflicts with minimization and increases risk. Legal counsel must confirm obligations for the deployment.", "references": [ "https://eur-lex.europa.eu/eli/reg/2016/679/oj", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-045", "domain": "governance-safety-risk", "objective": "Account for GDPR, HIPAA, FedRAMP, and applicable regulation", "type": "single", "prompt": "A healthcare workflow will send protected health information through Claude. What is the architect's strongest first governance action?", "options": [ "Assume encryption alone makes the workflow compliant", "Ask Claude to certify compliance", "Use the provider's encrypted transport and approved enterprise plan, send complete records to maximize clinical context, and rely on downstream access logs for compliance", "Confirm the covered workflow, approved service configuration and agreements, minimum-necessary data, access controls, auditability, and organizational HIPAA review" ], "correct": [ 3 ], "explanation": "HIPAA suitability depends on the complete administrative, technical, contractual, and data-minimization boundary. A model cannot declare the system compliant, and encryption is only one control.", "references": [ "https://www.hhs.gov/hipaa/for-professionals/security/index.html", "https://trust.anthropic.com/", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-046", "domain": "governance-safety-risk", "objective": "Account for GDPR, HIPAA, FedRAMP, and applicable regulation", "type": "single", "prompt": "A US government team requires a FedRAMP-authorized boundary. Which statement is correct?", "options": [ "The exact cloud service, deployment boundary, region, inherited controls, customer responsibilities, and data flow must be verified", "A local prototype is sufficient evidence", "Use the same model family through any government cloud region, provided encryption, identity federation, and agency logging meet the team's inherited control baseline", "FedRAMP applies only to the user interface" ], "correct": [ 0 ], "explanation": "Authorization attaches to a defined service and control boundary, not a model name in isolation. Architects must verify the approved offering and map inherited versus customer-operated controls.", "references": [ "https://www.fedramp.gov/", "https://trust.anthropic.com/", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-047", "domain": "governance-safety-risk", "objective": "Address bias, fairness, and transparency", "type": "multiple", "prompt": "A benefits-triage system has acceptable overall accuracy. Which fairness checks are still needed? Select all that apply.", "options": [ "Compare error and escalation rates across relevant groups", "Use the acceptable aggregate score plus a balanced overall label distribution as evidence of fairness, then monitor complaints for group-specific problems", "Test accessibility and language slices", "Review whether labels and proxy variables encode historical inequity" ], "correct": [ 0, 2, 3 ], "explanation": "Aggregate performance can conceal concentrated harm. Slice metrics, data and label review, and accessibility evaluation reveal distributional failures that require governance decisions.", "references": [ "https://www.nist.gov/itl/ai-risk-management-framework", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-048", "domain": "governance-safety-risk", "objective": "Address bias, fairness, and transparency", "type": "single", "prompt": "Users receive recommendations that affect account access. Which transparency design is strongest?", "options": [ "Publish private chain-of-thought", "Disclose the AI-assisted role, provide understandable factors and evidence where appropriate, and offer a challenge or escalation path", "Disclose that AI generated the recommendation, show a calibrated confidence band, and direct users to ordinary customer support without exposing decision factors", "Hide that AI contributed so users trust the result" ], "correct": [ 1 ], "explanation": "Meaningful transparency explains the system's role, basis, limitations, and recourse without exposing private reasoning or sensitive controls. A bare confidence number is not an explanation.", "references": [ "https://www.nist.gov/itl/ai-risk-management-framework", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-049", "domain": "governance-safety-risk", "objective": "Identify LLM failure modes", "type": "multiple", "prompt": "An external document instructs a procurement agent to approve a new vendor and send credentials. Which controls should prevent harm? Select all that apply.", "options": [ "Treat the document as authenticated business intent when its signature image and vendor details match an approved record, then apply ordinary tool approval", "Treat document instructions as untrusted content", "Require authenticated human approval bound to the exact vendor action", "Keep vendor approval and credential access outside automatic model authority" ], "correct": [ 1, 2, 3 ], "explanation": "Indirect prompt injection exploits legitimate agent authority. Trust labeling, least privilege, and action-bound approval prevent content from authenticating or authorizing itself.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/mitigate-jailbreaks", "https://modelcontextprotocol.io/docs/tutorials/security/security_best_practices", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-050", "domain": "governance-safety-risk", "objective": "Layer guardrails and safety controls", "type": "single", "prompt": "A secret appears in a model trace. What is the correct immediate response?", "options": [ "Restrict trace access, add the secret pattern to redaction, and monitor the credential for suspicious use before rotating it at the next maintenance window", "Ask the model to forget it", "Contain the affected capability, revoke or rotate the credential, preserve redacted evidence, and investigate actual access", "Delete the trace and continue using the credential" ], "correct": [ 2 ], "explanation": "Once exposed, a credential is no longer trustworthy. Containment and rotation reduce ongoing risk, while protected evidence and authoritative logs establish impact and support regression fixes.", "references": [ "https://code.claude.com/docs/en/security", "https://www.nist.gov/itl/ai-risk-management-framework", "certifications/claude/lessons/06-governance-safety-and-responsible-use" ] }, { "id": "ccar-p-m1-051", "domain": "stakeholder-lifecycle", "objective": "Run structured discovery", "type": "multiple", "prompt": "A discovery workshop for an underwriting assistant is beginning. Which questions should the architect prioritize? Select all that apply.", "options": [ "Which data and systems are authoritative", "Which action follows the output and what happens when it is wrong", "What observable outcome should improve and who owns it", "Which current model leads broad public benchmarks and has the strongest adoption among comparable underwriting teams" ], "correct": [ 0, 1, 2 ], "explanation": "Outcome ownership, downstream consequence, and authoritative data define the problem boundary. Model popularity is not a requirement and should not precede discovery.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-052", "domain": "stakeholder-lifecycle", "objective": "Communicate decisions and tradeoffs", "type": "single", "prompt": "Executives ask why the proposed architecture keeps human approval for refunds above a threshold. Which explanation is strongest?", "options": [ "The engineering team prefers more screens", "The model is not intelligent enough", "Human review is retained because refunds are external financial actions, so a uniform approval rule across all values is simpler to audit and explain than a calibrated risk threshold", "The threshold balances handling-time benefit with the financial impact, reversibility, policy obligation, and measured uncertainty of higher-value cases" ], "correct": [ 3 ], "explanation": "Tradeoff communication should connect the control to value, risk, evidence, and policy. It avoids vague capability claims and explains why the boundary is proportional.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://www.nist.gov/itl/ai-risk-management-framework", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-053", "domain": "stakeholder-lifecycle", "objective": "Manage feedback, expectations, and SLAs", "type": "single", "prompt": "A stakeholder says the agent must be 99 percent accurate. What should the architect ask next?", "options": [ "Which task population, label, evaluator, time window, error classes, and consequences define that target", "Whether 99 is the stakeholder's favorite number", "Which prompt is longest", "Whether the target refers to offline model accuracy or end-to-end production success, then adopt the stricter definition to avoid understating performance" ], "correct": [ 0 ], "explanation": "An accuracy target is meaningless without a denominator, measurement method, and error taxonomy. Operational commitments must name the population and response when objectives are missed.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-054", "domain": "stakeholder-lifecycle", "objective": "Document architectures and implementation guidance", "type": "multiple", "prompt": "Which sections belong in a useful architecture decision record? Select all that apply.", "options": [ "Options, decision, and rejected alternatives", "Consequences, verification, and reversal conditions", "The final architecture diagram plus implementation milestones and owners, with rejected options and reversal conditions retained in workshop notes", "Context, constraints, and evidence" ], "correct": [ 0, 1, 3 ], "explanation": "A decision record preserves why an option fit the evidence and how it will be validated or reversed. A diagram alone shows structure but not rationale or operational consequences.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-055", "domain": "stakeholder-lifecycle", "objective": "Support discovery, design, handoff, monitoring, and iteration", "type": "multiple", "prompt": "Before handing a Claude system from the build team to operations, which artifacts are required? Select all that apply.", "options": [ "Runbooks, escalation paths, and rollback procedure", "A signed build-team statement that the production-like demo passed its acceptance cases, with unresolved issues documented in the handoff meeting notes", "Service objectives, dashboards, and alert ownership", "Version inventory for model, prompt, tools, data, and policy" ], "correct": [ 0, 2, 3 ], "explanation": "Operations needs measurable service behavior, response procedures, ownership, and versioned dependencies. A successful demo does not define how to detect, contain, or recover from production failure.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://sre.google/sre-book/table-of-contents/", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-056", "domain": "stakeholder-lifecycle", "objective": "Manage feedback, expectations, and SLAs", "type": "multiple", "prompt": "A high-impact assistant is ready for rollout. Which pilot design is strongest? Select all that apply.", "options": [ "Grant production authority to a small expert-user cohort with a global kill switch, using rapid incident response instead of draft-only or read-only boundaries", "Collect structured reviewer feedback and failure traces", "Use a bounded population and read-only or draft-only authority", "Define success, safety gates, review load, and rollback before launch" ], "correct": [ 1, 2, 3 ], "explanation": "A bounded pilot tests the riskiest assumptions while limiting harm. Predeclared gates and structured evidence make the decision to expand or stop defensible.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-057", "domain": "stakeholder-lifecycle", "objective": "Support discovery, design, handoff, monitoring, and iteration", "type": "single", "prompt": "Reviewers complain that too many safe cases are escalated. What should the lifecycle owner do?", "options": [ "Raise the escalation confidence threshold immediately based on reviewer complaints, then compare queue volume and incident counts during the next release window", "Measure false-escalation slices, review the trigger and evidence quality, test a candidate change, and monitor safety guardrails", "Tell reviewers to work faster", "Remove human review entirely" ], "correct": [ 1 ], "explanation": "Feedback should become a classified, measurable improvement hypothesis. The team can reduce unnecessary review only while preserving the high-impact safety boundary.", "references": [ "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-058", "domain": "stakeholder-lifecycle", "objective": "Communicate decisions and tradeoffs", "type": "single", "prompt": "How should the same architecture change be communicated to an executive sponsor and an on-call engineer?", "options": [ "Give both audiences the architecture decision record and full sequence diagram so each receives the same authoritative details and can extract what matters", "Send both the complete source code only", "Explain business outcome, risk, and cost to the sponsor; give dependencies, signals, failure modes, and recovery steps to the engineer", "Hide tradeoffs from the sponsor" ], "correct": [ 2 ], "explanation": "Stakeholders need different levels of detail while the underlying decision remains consistent. Communication should support each audience's decisions and responsibilities.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-059", "domain": "stakeholder-lifecycle", "objective": "Support discovery, design, handoff, monitoring, and iteration", "type": "multiple", "prompt": "A source system is being retired. Which lifecycle actions are required for the Claude application? Select all that apply.", "options": [ "Migrate or remove data with retention and deletion controls", "Identify prompts, retrieval indexes, tools, evals, and runbooks that depend on it", "Re-run acceptance and failure tests before cutover", "Keep the old connector disabled but deployable as an emergency fallback, without including it in routine runbooks, tests, or dependency inventories" ], "correct": [ 0, 1, 2 ], "explanation": "Decommissioning is an architectural change across data, integration, evaluation, and operations. Undocumented stale access increases security risk and can silently reintroduce obsolete information.", "references": [ "https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542810%2FClaude+Certified+Architect+%E2%80%93+Professional+Exam+Guide.pdf", "https://www.nist.gov/itl/ai-risk-management-framework", "certifications/claude/lessons/22-business-discovery-requirements-and-slas" ] }, { "id": "ccar-p-m1-060", "domain": "developer-productivity-operations", "objective": "Configure Claude tools and team environments", "type": "multiple", "prompt": "A team standardizes Claude Code across repositories. Which configuration belongs in reviewed project scope? Select all that apply.", "options": [ "Repository-specific constraints and approved hooks", "Versioned team settings needed for consistent safe behavior", "References to production secret names together with sample resolved values so every repository can validate credential format during setup", "Canonical build and test commands" ], "correct": [ 0, 1, 3 ], "explanation": "Shared commands, constraints, hooks, and settings create repeatable behavior. Credentials belong in protected environment or secret-management boundaries, never committed project instructions.", "references": [ "https://code.claude.com/docs/en/settings", "https://code.claude.com/docs/en/hooks-guide", "certifications/claude/lessons/12-claude-agent-sdk-and-hooks" ] }, { "id": "ccar-p-m1-061", "domain": "developer-productivity-operations", "objective": "Improve engineering workflows with AI-assisted tooling", "type": "multiple", "prompt": "A headless Claude Code job reviews untrusted pull requests. Which controls are appropriate? Select all that apply.", "options": [ "Minimal repository and token permissions", "A short-lived protected-branch token limited to reviewer-generated fixes, allowing the headless job to apply high-confidence findings after tests pass", "Bounded tools, network, time, and structured report output", "No unrelated secrets in the environment" ], "correct": [ 0, 2, 3 ], "explanation": "Headless automation lacks interactive supervision, so capability and identity must be narrower. Findings should enter normal review and protected merge paths instead of bypassing them.", "references": [ "https://code.claude.com/docs/en/headless", "https://code.claude.com/docs/en/security", "certifications/claude/lessons/12-claude-agent-sdk-and-hooks" ] }, { "id": "ccar-p-m1-062", "domain": "developer-productivity-operations", "objective": "Support debugging and operational incident resolution", "type": "multiple", "prompt": "An on-call engineer uses Claude to investigate a latency incident. Which inputs and boundaries are most useful? Select all that apply.", "options": [ "Short-lived read credentials copied into a protected system prompt so Claude can reproduce failing dependency requests directly from the incident context", "Redacted correlated traces and recent version changes", "Read-only diagnostic tools before any remediation authority", "Service objectives, dependency health, and runbook steps" ], "correct": [ 1, 2, 3 ], "explanation": "Correlated evidence, objectives, dependency state, and safe diagnostic tools support fast causal analysis. Credentials add exposure and should remain inside trusted integrations.", "references": [ "https://opentelemetry.io/docs/specs/otel/", "https://sre.google/sre-book/table-of-contents/", "https://code.claude.com/docs/en/security", "certifications/claude/lessons/12-claude-agent-sdk-and-hooks" ] }, { "id": "ccar-p-m1-063", "domain": "developer-productivity-operations", "objective": "Configure Claude tools and team environments", "type": "multiple", "prompt": "A team upgrades Claude Code, a plugin, and an MCP server in one release. Which practices reduce operational risk? Select all that apply.", "options": [ "Run representative canary workflows with rollback instructions", "Inventory changes to tools, hooks, permissions, and network access", "Record and pin versions where supported", "Rely on successful installation and protocol negotiation as compatibility proof, then monitor production canaries instead of running representative workflows before release" ], "correct": [ 0, 1, 2 ], "explanation": "Behavior depends on the combined tool and configuration surface. Version records, capability review, canaries, and rollback turn upgrades into controlled operational changes.", "references": [ "https://code.claude.com/docs/en/settings", "https://modelcontextprotocol.io/docs/tutorials/security/security_best_practices", "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests", "certifications/claude/lessons/12-claude-agent-sdk-and-hooks" ] } ] }