CCAR-P Practice Questions: Solution Design & Architecture Domain

Correct answer (A): A sequential workflow directly represents the fixed order and explicit dependencies while preserving each approved artifact for audit and replay. The decisive constraints are reproducibility and the requirement that every stage consume a validated predecessor. The architectural principle is to use deterministic orchestration when control flow is known rather than introducing model-directed autonomy. In production, explicit stage contracts simplify testing, localize failures, and give auditors evidence of how each statement was produced.

Why the other options are wrong:
- Option B: An autonomous agent is appropriate when the next action must be discovered dynamically, but here the steps and their order are fixed and auditable.
- Option C: Specialist debate could improve some drafts, but it introduces coordination overhead and nondeterminism without addressing a need for independent parallel work.
- Option D: A single call is operationally simple, but it obscures intermediate validation and weakens reproducibility when every stage requires an approved artifact.

Correct answer (A): Claude adds value when interpreting varied unstructured layouts, while exact identifiers, arithmetic, and duplicate detection require deterministic enforcement. The explicit requirement that financial controls be exact makes this partition best. Separating probabilistic interpretation from hard controls reduces payment risk and makes failures easier to audit in production.

Why the other options are wrong:
- Option B: Confidence-based review is a plausible exception strategy, but model-reported confidence does not guarantee exact arithmetic, identifier matching, or duplicate prevention.
- Option C: Template-based OCR offers predictable processing, but the scenario states that layouts vary widely, making it likely to preserve excessive manual work.
- Option D: Independent agreement may reduce some extraction errors, but two probabilistic calls do not replace deterministic financial validation and add unnecessary cost.

Correct answer (D): Independent extraction tasks can run concurrently, but covenant analysis and final synthesis cannot begin reliably until their required artifacts exist and have been validated. The latency target favors parallelism only where the dependency graph permits it. CCAR-P architecture derives orchestration from actual data dependencies rather than applying one pattern uniformly. This reduces wall-clock time while preserving schema validation, analytical ordering, and fault isolation.

Why the other options are wrong:
- Option A: Broad parallelism can reduce latency, but risk analysis performed before normalized dependent inputs exist may use inconsistent or incomplete figures.
- Option B: Risk-based early stopping can save processing in screening workflows, but the required final memo depends on all verified findings and cannot omit lower-priority documents.
- Option C: A single request avoids explicit orchestration, but it obscures dependency validation, increases context pressure, and provides weaker stage-level retry and fault isolation.

Correct answer (B): The initiative lacks success criteria and has materially consequential errors, so the architect must first establish measurable outcomes, failure costs, and step-level automation boundaries. The governing principle is to define the business problem and risk model before selecting an AI pattern. In production, this prevents a technology-first prototype from silently defining policy, controls, and acceptable risk.

Why the other options are wrong:
- Option A: A prototype can provide useful evidence, but beginning with full autonomy assumes the solution boundary before acceptable outcomes and failure costs have been defined.
- Option C: A capable model and universal review may reduce some risk, but they do not establish which workflow steps require Claude or define measurable production success.
- Option D: Retrieval may support correspondence analysis, but it addresses data access before the team has determined whether retrieval or automation is appropriate for each step.

Correct answer (C): Most requests have a strict latency requirement, while only a smaller segment has a high cost of error and can tolerate review. Segmenting traffic preserves routine-response latency and reserves expensive analysis and oversight for cases where they create business value. Production architecture should optimize quality, latency, cost, and failure impact together rather than maximize one aggregate score.

Why the other options are wrong:
- Option A: The higher accuracy is attractive, but the six-second response violates the 1.5-second requirement for most traffic and triples cost without segment-specific justification.
- Option B: A uniform fast path meets latency, but post-resolution auditing is insufficient for costly fraud or termination errors that should be prevented before action.
- Option D: Asynchronous comparison can support evaluation, but it does not provide timely protection for high-risk cases and imposes multi-agent cost across all requests.

Correct answer (C): Only the intake workflow has validated demand, and the small operating team cannot justify the support burden of a broad autonomous platform. A modular implementation meets current needs while preserving reversible interfaces for later expansion. This approach values maintainability and cost per successful business task rather than paying immediately for speculative flexibility.

Why the other options are wrong:
- Option A: A general platform may reduce future migration effort, but its current monitoring, testing, and incident-response costs are unsupported by confirmed demand.
- Option B: Independent stacks can optimize specialized workloads, but building them for hypothetical use cases creates duplication and an excessive ownership burden for two engineers.
- Option D: A single agent may appear operationally simple, but unrestricted tools weaken boundaries and do not provide modular ownership as the use cases expand.

Correct answer (D): The work is tightly sequential, and measured failures originate in handoffs and repeated context reconstruction rather than insufficient specialization. A workflow using shared, versioned state and validated artifacts removes the root cause while retaining stage boundaries. Architectural simplicity improves cost, consistency, and diagnosability when autonomous coordination provides no demonstrated benefit.

Why the other options are wrong:
- Option A: A supervisor could detect some inconsistencies, but it adds another model-mediated handoff instead of removing the snapshot and context fragmentation causing defects.
- Option B: Shared repository visibility may reduce inconsistent assumptions, but duplicating the full context across agents increases the identified cost and retains peer-coordination complexity.
- Option C: Redundant workers and judging can improve candidate selection, but they multiply cost and do not establish consistent state across dependent stages.

Correct answer (A): A structured handoff must preserve provenance and uncertainty at the point where facts are extracted. The explicit requirement is that reviewers trace each recommendation to source evidence, which prose summaries cannot reliably support. The governing principle is to use durable, versioned contracts between probabilistic stages and reject incomplete artifacts rather than asking later stages to infer missing information. In production, this supports validation, replay, schema evolution, and accountable clinician review.

Why the other options are wrong:
- Option B: A consistency check may detect contradictions with the summary, but it cannot recover source provenance that the extraction stage failed to preserve.
- Option C: Longer summaries could retain more context, but they do not create explicit evidence links, uncertainty fields, or machine-verifiable completion states.
- Option D: Central logs aid investigation and audit, but they are a compensating control rather than a reliable source-linked contract between pipeline stages.

Correct answer (C): The workload contains distinct latency and risk classes that should not be forced into one service pattern. Routine, bounded interactions can satisfy the two-second expectation synchronously; long-running evidence analysis belongs in a retryable asynchronous workflow with visible status; and consequential denials require the mandated approval gate. Aligning service boundaries with latency tolerance and failure impact creates more business value than maximizing a single automation metric.

Why the other options are wrong:
- Option A: A single synchronous flow simplifies the interface, but multi-minute analysis cannot meet the two-second interaction target and is poorly suited to retries.
- Option B: Asynchronous execution fits complex analysis, but applying it to routine status requests unnecessarily violates the explicit rapid-response expectation.
- Option D: Specialization may improve modularity, but independent action omits the required approval for denials and does not distinguish synchronous from long-running processing.

Correct answer (D): The migration should be decomposed around architectural dependencies rather than file layout or unconstrained agent decisions. Shared contracts, deployment ordering, and owner review require explicit understanding and planning artifacts before implementation. Bounded work packages limit blast radius, while independent contract verification prevents the implementing model's assumptions from becoming the acceptance standard. This creates reviewable handoffs and preserves cross-service compatibility.

Why the other options are wrong:
- Option A: Directory partitioning enables concurrency, but it ignores the stated cross-service contracts and defers predictable dependency failures until costly integration.
- Option B: More context and smaller commits may improve local execution, but they do not create an explicit dependency plan or independent interface verification.
- Option C: Service specialization is plausible, but local-only context causes agents to miss the shared interfaces that the scenario identifies as the main failure source.