CCAR-P Mock 2 · Judgment — Free Practice (42 Questions)

P1 · Q1Best-next-action / solution scoping

A national hospital network wants to pilot a Claude-powered assistant that drafts discharge summaries from clinician notes. Legal has confirmed that protected health information will flow through the system, the CISO requires all processing to remain inside the organization's approved cloud tenancy, and the clinical safety board insists every generated summary be reviewed by a licensed clinician before it enters the patient record. The delivery team is under pressure to ship something in six weeks. What is the best next action for the solution designer?

P1 · Q2Decision-matrix: platform selection under residency

A European logistics carrier standardizes its cloud estate on a single hyperscaler for which a Claude managed offering exists, and its regulator requires that shipment and customs data never leave the EU. The product team also wants access to Claude's newest tool-use features as soon as they are released. Given that the managed platform offers in-region deployment but sometimes lags the direct API on new features, which platform decision best fits the constraints?

P1 · Q3Reference-architecture selection & switching cost

A public-sector benefits agency built its first Claude integration by hard-coding calls to a single model version behind a thin wrapper, and now wants to add a second use case with different latency and cost needs. Leadership is worried about being locked into today's design as models and requirements evolve. Which reference-architecture choice best reduces future switching cost without over-engineering the current pilot?

P1 · Q4Outcome-document / auditable controls

A retail bank's design authority will only approve a new Claude-based fraud-triage assistant if the solution design document demonstrates that each identified risk has a control that is actually in place and someone accountable for verifying it stays effective. A designer submits a document listing risks and mitigations. Which characteristic most needs to be present for the document to satisfy the design authority?

P1 · Q5Best trade-off: batch vs interactive

A supply-chain company must re-classify roughly 12 million historical shipment descriptions into a new taxonomy for an annual regulatory filing. The job has no user waiting on it, must finish within a day, and cost is the dominant concern for the finance team. The same Claude prompt would also power a real-time classification widget for warehouse staff. How should the designer split the workload for the best cost/latency trade-off?

P2 · Q6Prompt caching on stable prefix

A gov-services contact center runs a Claude assistant where every request begins with the same 30,000-token block of policy guidance and eligibility rules, followed by a short, caller-specific question. Traffic is heavy and steady during business hours, and the team wants to cut token cost and time-to-first-token without changing answer quality. Which design change best achieves this?

P2 · Q7tool_choice semantics

A retailer's Claude-based shopping assistant has a get_inventory tool. Product managers complain that for simple greetings and general questions the assistant sometimes calls the inventory tool unnecessarily, and they want the model to decide for itself when a lookup is actually warranted rather than being forced. The team is reviewing the tool_choice setting. Which configuration matches the intended behavior?

P2 · Q8RAG vs long-context

A hospital group is building a clinical-guidelines assistant over about 8,000 guideline documents that are revised continuously as evidence changes. Clinicians must see the exact source passage cited, answers must never quote a superseded guideline, and responses must return within a couple of seconds. Engineers debate whether to place the whole corpus in Claude's long context each call or to add retrieval. Which approach best fits the constraints?

P2 · Q9Model routing under SLA

A logistics control tower uses Claude for two paths: a high-volume path that extracts structured fields from carrier emails under a tight latency SLA, and a low-volume path that writes nuanced exception-handling recommendations for complex disruptions where reasoning quality matters most. Budget is finite. How should the designer route between a faster, cheaper model and a more capable model?

P2 · Q10Context strategy: caching vs retrieval decision

A public records office runs a Claude assistant whose prompt contains two parts: a 20,000-token standard operating procedure that almost never changes, and a citizen-supplied case file that is different for every request and can be large. The team wants lower cost and faster responses while keeping full fidelity on both parts. Which context strategy is most appropriate?

P3 · Q11MCP vs direct integration

A hospital IT team wants Claude to reach several internal systems — an EHR lookup service, a bed-management API, and a formulary database — and expects to add more systems over the next year. Different teams own each system, and the group wants a consistent, governable way to expose tools to Claude without rebuilding the integration glue each time. Which integration approach best fits?

P3 · Q12Decision-matrix: platform standardization vs feature access

A retail group already runs its entire data platform on one hyperscaler and mandates that all AI workloads use that hyperscaler's managed services for billing and governance consistency. A new merchandising team wants to prototype with a just-released Claude capability that is currently only on the direct Anthropic API and not yet on the managed platform. There is no data-residency or regulatory constraint on this prototype. What is the best-fit decision?

P3 · Q13Least-risky integration: platform failover

A logistics platform serves a Claude-powered routing assistant to dispatchers under a strict availability SLA. It runs primarily on an in-region managed cloud platform for data residency. Leadership asks for a failover option so a regional outage of that platform does not take the assistant fully offline. Which failover design is least risky given the residency requirement?

P3 · Q14Integration correctness: tool result handling and grounding

A public-sector permitting portal integrates Claude with a live fee-calculation microservice via a tool. During testing, when the tool returns an error or empty result, the assistant sometimes states a plausible-sounding fee anyway. Officials require that any fee shown to an applicant come only from the authoritative service. What is the most appropriate integration fix?

P4 · Q15LLM-as-judge vs human evaluation

A healthcare provider is building an offline eval harness for a Claude-based agent that drafts patient-visit summaries from clinician notes. The team needs to score 8,000 draft summaries per release for clinical faithfulness (no hallucinated diagnoses or omitted allergies). They have three staff clinicians available roughly two hours each per week, and want repeatable scores that gate the release. What is the best evaluation design?

P5 · Q16Fail-closed vs fail-open guardrails

A logistics company runs a Claude agent that can call a tool to release freight from bonded customs warehouses. An output guardrail checks each release instruction against sanctions and license rules. During a deploy, the guardrail service starts timing out on ~15% of calls. The on-call engineer must decide the default behavior when the guardrail cannot return a verdict in time.

P5 · Q17Signal triage / escalation triggers

A public-sector benefits agency runs a Claude assistant that helps caseworkers draft eligibility determinations. Governance defined a pre-agreed monthly stakeholder review and a rule that any month where the assistant's projected inference spend exceeds the approved quarterly budget line triggers an immediate stakeholder escalation. In week 2, projected spend crosses that budget line. A separate dashboard also shows the model's average self-reported confidence dipped from 0.94 to 0.92. Which signal should escalate now, and why?

P4 · Q18Schema-valid JSON is not correct output

A retail company's Claude agent extracts structured product attributes (dimensions, material, hazardous-flag) from supplier PDFs into a JSON schema that feeds the catalog. The team added strict JSON-schema validation and every response now passes. A merchandiser reports that several dangerous-goods items are still shipping mislabeled. What is the best next step to improve output quality?

P5 · Q19Compliance control register

A hospital network must demonstrate to auditors that its Claude-based clinical-coding assistant complies with a regulation requiring that any automated code suggestion affecting billing is traceable to a human sign-off. The compliance lead asks how to make this obligation auditable. What is the best approach?

P4 · Q20Calibrated threshold vs raw confidence

A supply-chain company routes Claude's damage-claim assessments to auto-approve or human-review. Currently any claim where the model outputs a self-reported confidence at or above 0.92 is auto-approved. An audit finds auto-approved claims have a higher error rate than the manual queue. The data-science lead wants a principled routing rule. What should they do?

P5 · Q21Guardrail placement (input vs output vs tool-call authorization)

A government tax portal deploys a Claude agent that can call a tool to issue refund payments. The security team wants to prevent both prompt-injection attempts embedded in taxpayer correspondence and unauthorized payment actions. They currently have only a single output filter that scans the model's final text reply. How should guardrails be placed?

P4 · Q22Eval as acceptance criteria

A retail team is one week from launching a Claude-powered returns assistant. Leadership asks whether it is 'good enough to ship.' The team has a demo that works well on a handful of hand-picked conversations but no defined pass bar. What should they do before committing to launch?

P5 · Q23Alignment boundary / rule never given

A logistics operator's Claude agent handles driver-support chats. A post-incident review finds the agent disclosed another driver's home address when a caller claimed to be that driver's dispatcher. The team's system prompt covered tone and escalation paths but never stated a rule about verifying identity before releasing personal data. What is the most accurate characterization and best fix?

P4 · Q24Cost/latency optimization with Batches

A healthcare analytics team reprocesses 400,000 archived clinical documents nightly to update a de-identification index. The job is not time-sensitive and can complete any time before the next morning. They currently send every document through the standard synchronous API and costs are high. What is the best optimization?

P5 · Q25Signal triage / regulated checkpoint on schedule

A public-sector agency's Claude document-triage system is governed by a regulatory requirement for a quarterly independent model review, with the next review due in six weeks. Midway through the quarter, internal monitoring shows output quality is stable and well within tolerance. An engineer proposes skipping the upcoming quarterly review since the metrics look healthy. How should this be handled?

P4 · Q26Signal triage / review tier routing

A retail company monitors its Claude shopping assistant in production. Over one day it observes three signals: (a) a single customer tweet calling one reply 'a bit robotic'; (b) internal latency p95 rising from 2.1s to 2.4s, still within SLA; (c) a newly discovered pattern where the assistant recommends out-of-stock items 8% of the time, above the agreed 2% correctness threshold. Which signal warrants escalation beyond routine monitoring, and to where?

P5 · Q27Fail-closed control on a degrading dependency

A hospital's Claude medication-interaction assistant depends on an external drug-database API to validate its suggestions before they reach a pharmacist. The API begins returning stale data flagged with an 'unverified' status. The team must decide how the assistant behaves when validation comes back unverified.

P5 · Q28Single-output-filter anti-pattern

A logistics firm's customer-facing Claude agent both answers questions and can call tools to reschedule deliveries and issue account credits. Security review finds the only safety mechanism is one filter on the model's final natural-language reply. Leadership wants to know the biggest weakness in this design. What is the most accurate assessment?

P6 · Q29Recommend-the-option: missing reversal cost in sponsor presentation

A regional hospital network is deciding how to roll out an AI clinical-documentation assistant to 400 clinicians. Your team prepares a one-page option summary for the executive sponsor. Option A (staged pilot) reads: 'Gains: 30% faster note completion for the pilot ward; Gives-up: two extra weeks before full launch.' The sponsor asks whether the presentation is complete enough to approve. Which single element is missing that would make Option A a sound recommendation?

P6 · Q30Outcome document: audit-ready fields and the proxy-metric trap

A national parcel-logistics carrier deployed an AI routing optimizer and must file a quarterly outcome document that internal audit will review. A draft reports: 'The optimizer reduced average planned route distance by 12%, proving operational improvement.' Audit will accept the document only if each claimed outcome is backed by an auditable control-in-place and a named measurement owner. Which revision makes the document audit-ready?

P7 · Q31Review discipline: verification checklist for AI-generated code

A retail company's platform team lets engineers use an AI coding assistant. A junior engineer submits a PR for a checkout discount service; the AI-generated code passes all existing unit tests and the JSON config validates against the schema. The tech lead is under pressure to merge before a promotion launch. What should the team's review discipline require before merge?

P6 · Q32Handoff: a design only the architect understands

In a state unemployment-benefits agency, a lead architect built an AI eligibility-triage pipeline with several custom orchestration steps. The architect is rotating off to another program. During handoff the receiving team can run the system but cannot explain why certain cases are routed to manual review, and the routing rules live only in the architect's head and a private notebook. What is the most important action before the architect departs?

P7 · Q33Ops symptom to architecture cause: subagent context

A logistics SaaS provider runs a multi-agent system: a coordinator agent gathers a shipment's constraints, then spawns subagents to book carriers. Operations reports that subagents frequently book carriers that violate constraints the coordinator clearly collected, even though each subagent's own logs look internally consistent. What architectural cause best explains the symptom?

P7 · Q34Skills distribution: versioning and rollout ownership

A hospital IT group built an internal AI 'Skill' that formats discharge summaries to the organization's template. Different departments now run slightly different copies they each tweaked, and a recent template change caused inconsistent outputs that reached patients. Leadership asks how to prevent recurrence while keeping the Skill available org-wide. What is the core gap to fix?

P6 · Q35When a rollout needs re-approval

A retail chain received governance approval to pilot an AI product-recommendation model on its website for logged-in loyalty members only. The pilot succeeds, and the growth team wants to extend it to all anonymous visitors and also feed it into the email marketing channel next week, reusing the existing approval to move fast. What should you advise?

P6 · Q36Recommend-the-option among Gains/Gives-up/Reversal summaries

A municipal transit authority must pick how to deploy an AI incident-summarization tool for its control room. Option A: buy a vendor SaaS — Gains: live in 3 weeks; Gives-up: incident data leaves the authority's tenancy; Reversal cost: low, cancel the contract. Option B: self-host open weights — Gains: data stays in-tenancy; Gives-up: 3 months to production; Reversal cost: high, sunk infra build. The sponsor's stated top constraint is a legal mandate that resident incident data must not leave government systems. Which should you recommend, and why?

P7 · Q37Skills distribution + spend controls

A supply-chain analytics firm gave every analyst access to an AI assistant with a shared Skill that queries a paid third-party freight-rate API on each call. Within a month the API bill tripled and finance flagged it. Leadership wants analysts to keep the capability without runaway cost. What is the most effective control to add?

P6 · Q38Which stakeholder to involve first

A pharmacy retailer is building an AI tool that suggests generic-substitution options to pharmacists at the counter. Engineering has a working prototype and wants to schedule a launch. Marketing, legal/regulatory, the pharmacist union, and the data-platform team all need to be looped in eventually. Given the domain, which stakeholder should be engaged first to avoid a late, expensive redesign?

P7 · Q39Team self-sufficiency vs. escalating on policy

A federal benefits contact center uses an AI assistant to draft responses to citizen inquiries. Agents increasingly accept the AI's answer whenever it states it is 'highly confident,' including on eligibility edge cases the policy manual addresses differently. Quality review finds several confidently-worded but policy-incorrect responses went out. What should the enablement approach change?

P6 · Q40Name-the-missing-element: compliance posture

A health-insurance startup pitches its sponsor to launch an AI prior-authorization assistant. The one-page recommendation lists: 'Gains: median decision time drops from 4 days to 6 hours; Gives-up: two engineers reassigned for a month; Reversal cost: moderate, revert to the manual queue within a day.' The sponsor, who answers to a regulator, asks what is missing before approval. Which single element is absent?

P7 · Q41Ops symptom to architecture cause: schema-valid not correct

A retailer's AI agent extracts supplier invoice data into a structured record that downstream systems consume. Operations reports that records pass schema validation and flow through without errors, yet the finance team keeps finding invoices posted to the wrong cost center. What is the most likely cause?

P6 · Q42Lifecycle/GTM: sign-off gate before general availability

A logistics platform ran a limited-availability launch of an AI dock-scheduling feature with three warehouse customers. Metrics are strong and sales wants to flip it to general availability at next week's industry conference for maximum publicity. The success metric tracked during the limited launch was 'appointments booked through the feature,' and support has not yet been staffed or trained for broad rollout. What is the soundest lifecycle recommendation?