CCAR-P Practice Questions: Claude Models, Prompting & Context Engineering Domain

Correct answer (C): The required evidence is already available, so the failure arises from context composition and dilution rather than missing retrieval. Ranking by relevance, authority, freshness, and diversity prevents repetitive older sources from dominating; progressive expansion limits latency, and retained provenance supports traceability. CCAR-P context engineering treats context capacity as a budget to curate, not an instruction to include everything, because excessive competing evidence can reduce grounded synthesis quality.

Why the other options are wrong:
- Option A: Priority instructions may influence model behavior, but leaving 160 repetitive and conflicting documents in context preserves the dilution causing the failure.
- Option B: A larger context window can support broader research, but adding more evidence worsens the identified competition among sources and may violate the latency SLA.
- Option D: Summaries could lower token use, but removing links to original evidence violates the claim-traceability requirement and introduces another lossy interpretation layer.

Correct answer (C): Ranking and deduplication remove low-value competition for attention, while provenance-preserving treatment of conflicts satisfies researchers' inspection requirement. Progressive loading makes secondary evidence available without overwhelming the initial synthesis. The governing principle is to engineer context for relevance, salience, and traceability rather than fill the available window. This improves grounded synthesis even when all material technically fits.

Why the other options are wrong:
- Option A: Repeating regulator notices may improve their salience, but retaining duplicates and weakly related sources continues to dilute evidence and consume attention.
- Option B: Pre-summarization can reduce context size, but removing attribution violates the provenance requirement and can conceal meaningful disagreements.
- Option D: Parallel analysis can support independent research tasks, but random partitioning separates related evidence and majority voting can suppress authoritative minority sources.

Correct answer (B): The shared guide and schema are substantial stable content, but request-dependent interleaving prevents effective prefix reuse. A consistently ordered, versioned stable region directly addresses the low cache-hit rate while isolating volatile customer data and contracts. The principle is to design prompts around stability boundaries so caching can reduce repeated token processing and latency without changing model behavior or output quality.

Why the other options are wrong:
- Option A: Caching can benefit shared material, but customer contracts are request-specific and may also create privacy and invalidation problems if placed in a common cached region.
- Option C: A different model could affect quality or instruction needs, but output quality is already acceptable and model capability does not fix unstable prompt composition.
- Option D: Summarization may reduce tokens, but doing it per request adds latency and cost while failing to reuse the unchanged policy guide.

Correct answer (B): The faster model is the best choice because it exceeds the explicit 95% macro F1 requirement and satisfies the 800 ms SLA, whereas the more capable model violates latency and adds substantial cost for a small quality gain. The governing principle is to select the least costly model configuration that reliably meets measured production requirements. This prevents unnecessary inference expense and preserves service performance at high volume.

Why the other options are wrong:
- Option A: A larger quality margin can be valuable for high-risk tasks, but this workload already exceeds its stated quality target, and the more capable model violates the 800 ms SLA.
- Option C: Traffic splitting can support experiments, but operating both models indefinitely does not provide a clear architectural benefit and sends some requests through a path that violates the SLA.
- Option D: Asynchronous processing can suit noninteractive workloads, but it changes the service behavior rather than meeting the stated response requirement and retains the fourfold cost increase.

Correct answer (D): The evidence shows two distinct problems: migration completeness and machine-readable structure. Staging makes the migration plan and rollback requirements inspectable before artifact generation, while deterministic schema and semantic checks prevent unsafe artifacts from entering the payroll pipeline. The governing principle is that reasoning quality and format compliance require different controls; production automation should reject, repair, or escalate failures rather than relying on imperative wording or unvalidated model judgment.

Why the other options are wrong:
- Option A: Clearer instructions may improve compliance, but repeated imperative wording cannot deterministically guarantee either valid structure or complete rollback logic.
- Option B: Generating alternatives can improve selection in some creative tasks, but a model-based score does not provide deterministic validation for payroll-affecting deployment artifacts.
- Option C: Additional syntax examples may reduce malformed manifests, but automatically accepting every output leaves missing rollback semantics and residual structural errors uncontrolled.

Correct answer (C): The logs show that the correct source is retrieved but diluted by numerous obsolete, similarly worded passages. Authority-aware reranking and filtering directly address context selection and salience, while clear context structure reduces instruction-source confusion. The governing principle is to fix the failing context layer before increasing model capability or context volume; otherwise production cost rises while conflicting evidence remains.

Why the other options are wrong:
- Option A: A stronger model may reconcile difficult evidence more effectively, but the known failure is poor evidence composition, and retaining all conflicting passages leaves that cause intact.
- Option B: Prompt emphasis might make the exception process more salient, but it does not remove obsolete evidence or reliably identify which source is authoritative.
- Option D: Summarization can reduce token volume, but retrieving more low-quality material increases the chance that obsolete guidance contaminates the evidence summary.

Correct answer (B): The evidence shows that retrieval and clause identification are working; the failure concerns mapping ambiguous language to a known output convention. A compact set of representative mappings demonstrates the desired treatment of edge cases without consuming excessive context. Few-shot prompting is most valuable when it targets the observed ambiguity and remains paired with explicit structured-output requirements.

Why the other options are wrong:
- Option A: Full contracts could expose additional language patterns, but they add substantial irrelevant context and latency when the failure is limited to a small set of known mappings.
- Option C: Additional reasoning instructions may be useful for comprehension failures, but the model already finds the correct clause and lacks a consistent mapping convention.
- Option D: Downstream normalization is technically possible, but removing the constrained enum broadens output variability and moves a prompt-level convention problem into more complex application logic.

Correct answer (A): Validated compaction distinguishes durable state from obsolete conversation content. It preserves the explicitly required decisions, tasks, and constraints while removing the stale tool output causing incorrect behavior. The governing principle is that context should be curated for relevance and currency rather than replayed indiscriminately. Production context lifecycle management reduces cost and prevents superseded information from competing with authoritative state.

Why the other options are wrong:
- Option B: Recency emphasis might improve salience, but retaining obsolete tool results preserves the conflict and continued token growth.
- Option C: Threshold-based truncation controls size, but indiscriminate removal can delete the approved decisions and constraints the system must retain.
- Option D: New sessions eliminate accumulated context, but repeatedly asking users to reconstruct state is unreliable and undermines the multi-day workflow.

Correct answer (C): The current request is underspecified, so explicitly defining mandatory analysis dimensions and a structured output contract directly addresses omissions and enables automated comparison. The architectural principle is to establish concrete success criteria before adding capacity or rhetorical emphasis. Clear contracts improve consistency and make production outputs testable.

Why the other options are wrong:
- Option A: More output capacity may help if responses are truncated, but the scenario identifies omitted requirements caused by an ambiguous prompt rather than token exhaustion.
- Option B: A stronger model can help with capability limitations, but no evaluation evidence shows that model capacity is the cause of these predictable omissions.
- Option D: Role prompting may affect style or perspective, but it does not specify the missing rollback, dependency, and testing requirements.

Correct answer (A): The large shared content is reusable, but volatile fields placed before it prevent the application from presenting a stable prefix. Moving the versioned manual and shared instructions ahead of per-request data directly targets the cause of poor cache reuse. Stable ordering and serialization reduce repeated processing cost and latency across requests with common context.

Why the other options are wrong:
- Option B: Longer retention can help when identical reusable content is recognized, but it does not fix the early request-specific fields that make the prompt prefix vary.
- Option C: Summarization could reduce token volume, but performing it for every request adds repeated processing and does not exploit the manual's monthly stability.
- Option D: Moving content between message roles does not by itself create a stable reusable prefix, and randomized ordering would further reduce consistency.