CCAR-P Practice Questions: Integration Domain

Correct answer (C): Direct exposure avoids discovery latency for the four dominant tools, while progressive discovery reduces context consumption and selection errors for the large changing catalog. Enforcing permissions during discovery prevents unauthorized capabilities from being offered. This hybrid approach applies architectural complexity only where catalog scale justifies it and preserves fast common-path performance.

Why the other options are wrong:
- Option A: Prompt caching may reduce repeated input processing, but it does not reduce the model-visible catalog or deterministically enforce user authorization.
- Option B: Universal discovery reduces initial context, but it adds unnecessary latency to common requests and a week-long cache may retain stale tools or permissions.
- Option D: Specialized agents can help with distinct complex domains, but extra routing and autonomy add latency and do not directly minimize exposed tool definitions.

Correct answer (D): Distributed trace correlation identifies which integration stage causes tail latency while structured fields support fleet-wide aggregation. Privacy-aware sampling addresses the restriction on transcript retention. This follows the production principle that model, retrieval, and tool behavior must be observable as separate but correlated components rather than diagnosed from averages or manual transcript review.

Why the other options are wrong:
- Option A: Transcripts can help inspect individual cases, but broad retention conflicts with the privacy constraint and does not efficiently isolate component latency at scale.
- Option B: Average latency is inexpensive to monitor, but it can hide tail behavior and cannot identify the responsible integration layer.
- Option C: Retries may mitigate transient failures, but indiscriminate retries can increase latency and do not diagnose the measured bottleneck.

Correct answer (B): Correlated stage-level traces distinguish retrieval, model, and tool latency while respecting the restriction on transcript storage. Percentile-compatible measurements, retry counts, and tool outcomes expose tail behavior that averages conceal. The production principle is to instrument integration boundaries with privacy-minimized telemetry so operators can localize failures without collecting unnecessary sensitive content.

Why the other options are wrong:
- Option A: Full transcripts can help inspect individual failures, but routine storage violates the privacy constraint and does not provide scalable component attribution.
- Option C: Aggregate dashboards aid capacity and cost monitoring, but averages cannot identify which stage causes p95 regressions.
- Option D: Retries may help some transient failures, but they can worsen tail latency and are a mitigation applied before locating the cause.

Correct answer (B): Security trimming must occur before restricted content enters model context and must cover semantic search, lexical search, reranking, and caching. Preserving access metadata and provenance supports both authorization and traceable citations. This structural boundary prevents disclosure rather than relying on the model or an output filter after unauthorized retrieval has already occurred.

Why the other options are wrong:
- Option A: Output redaction can be a compensating control, but restricted content has already entered context and may influence or leak through the answer.
- Option C: Index partitioning can support classification boundaries, but delegating project-level authorization to Claude leaves a required permission dimension unenforced.
- Option D: Encryption protects data at rest, but a retrieval service holding the key can still expose decrypted restricted chunks to unauthorized model requests.

Correct answer (A): Exact identifiers and minute-by-minute inventory values belong in authorized structured lookups, while conceptual document questions benefit from hybrid lexical and semantic retrieval with reranking. Preserving provenance satisfies the citation requirement. Matching retrieval strategy to data shape and freshness avoids forcing one mechanism to handle fundamentally different workloads.

Why the other options are wrong:
- Option B: A combined index is simple, but nightly inventory exports cannot satisfy minute-level freshness, and larger top-k does not ensure exact identifier matching.
- Option C: Keyword search handles codes well, but using it exclusively reduces recall for conceptual troubleshooting questions that semantic retrieval already handles effectively.
- Option D: Prompt examples may improve formatting, but semantic search remains unreliable for similar exact IDs, and periodic embedding refreshes still lag live inventory.

Correct answer (B): The agent needs autonomous diagnostic access but must not make unapproved production changes. Separating narrow read-only tools from approval-gated state-changing capabilities enforces that boundary structurally. The governing principle is least privilege combined with capability separation, which prevents destructive action rather than merely detecting it after execution. This matters in production because logging cannot reverse an outage or data loss caused by an overprivileged agent.

Why the other options are wrong:
- Option A: An architect might add real-time alerts as a rapid compensating control, but alerts do not prevent the agent from executing a destructive command before an operator responds.
- Option C: Prompt restrictions and command deny lists can reduce common mistakes, but the unrestricted shell still permits unanticipated state-changing commands and prompt-level enforcement is not deterministic.
- Option D: Isolation and audit review can limit or reconstruct some damage, but the shared credential still allows the agent to change production state without the required approval.

Correct answer (B): The trace identifies three independent searches as the measured bottleneck, so they can run concurrently without changing semantics. Regulatory lookup must remain after entity extraction because it has a genuine data dependency, and cancellation limits unnecessary cost. This is dependency-aware optimization rather than indiscriminate parallelism.

Why the other options are wrong:
- Option A: Full speculative parallelism could reduce wall-clock time, but regulatory queries lack required company names and would generate the substantial charges the scenario prohibits.
- Option C: Longer timeouts and retries may improve completion rates, but they increase latency and do not address the sequential independent searches shown by traces.
- Option D: Multiple agents may support genuinely independent analysis, but they add coordination, calls, and nondeterminism to a workflow with known dependencies.

Correct answer (D): Restricted content must be excluded before it reaches model context, and citation requirements demand versioned provenance from each retrieved chunk through the generated response. Trusted scope filtering enforces the privacy boundary structurally, while metadata and audit records make claims traceable. In production, post-generation filtering cannot reliably undo information already exposed to the model.

Why the other options are wrong:
- Option A: Encryption protects data while unreadable, and a policy service could control displayed citations. However, Claude cannot use encrypted chunks, while decrypting them into context would recreate the same pre-retrieval authorization failure.
- Option B: A second classifier can detect some leakage and regeneration may improve citation coverage. However, this remains a probabilistic compensating control after unauthorized content has entered context.
- Option C: Per-employee indexes can enforce strong isolation and may suit a small, static population. However, global scale and changing project permissions make document duplication and continuous rebuilding operationally excessive.

Correct answer (C): Selective escalation concentrates additional capability and cost on the small segment with the stringent quality requirement. The architecture must also evaluate missed high-risk cases because routing quality governs whether escalation works. This risk-based segmentation balances correctness, latency, and cost more effectively than applying expensive processing to the entire workload.

Why the other options are wrong:
- Option A: A universal upgrade would improve the high-risk segment, but it violates the stated cost and latency constraints, while reducing logging weakens operations.
- Option B: A repeated call may help with stochastic errors, but no evidence shows that the lower-cost configuration can reach the required high-risk accuracy.
- Option D: Ensembling is technically viable for difficult cases, but three calls on every ticket add unnecessary latency and cost to the low-risk majority.

Correct answer (C): Progressive discovery reduces context consumption and selection ambiguity because each request needs only a few of 240 tools. Filtering discovery by authorization prevents unavailable capabilities from being exposed, while measuring discovery latency protects the four-second SLA. The design accepts added routing complexity only where the catalog's scale and rate of change justify it.

Why the other options are wrong:
- Option A: Shorter descriptions may reduce context use, but all 240 tools remain exposed and prompt instructions do not enforce user authorization.
- Option B: Static grouping can reduce catalog size, but department-level grants are too coarse for the stated user-specific permissions and changing catalog.
- Option D: Platform-specific models could narrow selection, but training and maintaining multiple models is unnecessary and still exposes complete platform catalogs.