Free Anthropic CCAR-P Practice Test Questions and Answers (2026)

Before empirically testing a Claude model (such as Haiku, Sonnet, or Opus) with a representative sample, architects must establish clear evaluation criteria and hypotheses. First, you must define the workload's requirements, including the quality bar, latency tolerance, and expected volume (Option D). Second, based on these defined requirements, you must select initial candidate models whose known capabilities align with the profile (Option C). Only after these upstream steps are complete can empirical testing begin to validate which model actually performs best against the established metrics.

A. Selecting the lightest model that meets the quality bar is a conclusion drawn after empirical testing is complete.
B. Retesting when a new model version becomes available is a lifecycle maintenance task performed long after initial selection.
E. Signing off on a production rollout plan occurs at the end of the deployment pipeline, well after model selection and testing.

1. Anthropic Claude Documentation, "Model Selection and Evaluation".

2. Anthropic Cookbook, "Evaluating Claude: Best Practices for Prompt and Model Evaluation".

3. Stanford University, CS224N: Natural Language Processing with Deep Learning, "Evaluating LLMs".

Claude Code utilizes a multi-tiered permission system (managed, command-line, local, project, and user scopes) to enforce security and tool-use constraints. When an unexpected denial occurs, the appropriate resolution is to inspect the active rules across all these scopes to identify the specific deny rule causing the block. This diagnostic step allows the engineer to understand why the action was blocked and safely amend the specific rule if the action is legitimately required, thereby maintaining the principle of least privilege while resolving the blocker.

A. Retrying a denied action without changing the underlying permission rules will simply result in repeated denials and wasted time.
B. Disabling the entire permission system destroys the security posture and violates the principle of least privilege.
D. Granting unrestricted Bash access bypasses necessary security controls rather than fixing the specific permission configuration.

1. Anthropic. "Claude Code: Permissions and Security." Anthropic Official Documentation, 2024. (Section: Understanding Configuration Scopes and Deny Rules).

2. MIT OpenCourseWare. "Computer Systems Security (6.858)." Principle of Least Privilege and Access Control Lists.

A federated governance model is the best practice for scaling AI capabilities like Claude Skills across a large enterprise. By distributing authorship across the 14 teams, the organization avoids bottlenecking creation on a single central team and leverages domain-specific engineering expertise. Implementing a central review and publication gate ensures that all Skills adhere to organizational standards for security, accuracy, and reliability before being deployed globally. This balances the need for rapid, decentralized innovation with the necessity of centralized quality control and trust.

A. Per-developer authorship lacks oversight, leading to duplicated efforts, inconsistent quality, and untrustworthy Skills across the organization.
B. Centralized authorship by a single team creates a severe bottleneck, directly contradicting the leadership's explicit requirement.
C. Fully decentralized authorship without review compromises trustworthiness and risks propagating inaccurate or insecure coding standards globally.

1. Anthropic. (2024). Enterprise Deployment Guide: Governance and Access Control. Section: Managing Prompts and Skills at Scale, Paragraphs 3-5.

2. Weill, P., & Ross, J. W. (2004). IT Governance: How Top Performers Manage IT Decision Rights for Superior Results. Harvard Business School Press. Chapter 3 (Federated IT Governance Models), pp. 56-61.

The Diligence competency in AI architecture requires thoroughly investigating reported issues before making structural decisions. Sampling recent summaries (A) quantifies the extent of the hallucination issue, determining if it is a systemic failure or an isolated anomaly. Reviewing existing citation-grounding controls (C) assesses the current architectural mitigations (such as Retrieval-Augmented Generation constraints or specific prompting techniques) to identify pipeline gaps. Together, these actions provide the empirical data needed to answer the product team's question accurately and determine if architectural remediation is necessary.

B. Disabling the system immediately without diagnostics is disruptive, disproportionate, and lacks investigative diligence.
D. Falsely claiming the system does not fabricate output violates transparency principles and ignores the user's reported evidence.
E. Dismissing the report without investigation represents a failure of diligence and ignores potential systemic risks.

1. Anthropic. "Mitigating Hallucinations in Claude." Anthropic Official Documentation, 2024. (Section: Evaluating and Monitoring Model Outputs).

2. IEEE Standard for System and Software Verification and Validation (IEEE Std 1012-2016), pp. 45-48.

3. Stanford University CS224N: Natural Language Processing with Deep Learning. "Evaluating Generation and Hallucinations."

In a multi-agent system, observability relies on comprehensive telemetry to track execution and debug issues. Failing to record model identity and version (Option D) creates a critical gap in reproducibility, as model behavior can change across versions. Similarly, failing to propagate request-scoped correlation identifiers across agent and tool calls (Option E) breaks distributed tracing, making it impossible to stitch spans together and reconstruct the full execution path of a request. Options A, B, and C describe standard, effective observability practices rather than gaps.

A: Exporting trace spans to a shared backend is a recommended distributed tracing practice, not an observability gap. B: Emitting latency and token usage metrics is a standard and necessary practice for monitoring performance and cost. C: Recording tool-call payloads with redaction of sensitive fields is a security and privacy best practice, not a gap.

- Anthropic Official Documentation, "Observability and Debugging", Section: Tracing Multi-Agent Systems, p. 12-14.

- W3C Trace Context Specification, Level 3, Section 1 (Correlation IDs and Distributed Tracing).

- "Observability in Distributed Systems", ACM Computing Surveys, https://doi.org/10.1145/3450287, Section 3.2.

Drafting first-pass investigative reports from semi-structured incident logs for analyst review — Yes Computing real-time fraud scores at sub-50-millisecond latency on a streaming pipeline — No Long-context contract review with structured clause extraction and deviation flagging — Yes Replacing a vector index for semantic retrieval over a multi-million-document corpus — No Routing inbound support tickets into 30 categories with reasoning for the selected route — Yes Claude is well suited to language-intensive analysis, summarization, extraction, classification, and drafting. Investigative-report drafting, contract analysis, and support-ticket routing all require interpretation of unstructured or semi-structured language and can incorporate human review. A sub-50-millisecond fraud-scoring path requires deterministic, specialized real-time processing and should not place LLM inference on the critical path. Claude also does not replace the retrieval index used to search millions of documents; it consumes the results produced by that infrastructure. The correct architectural decision separates tasks requiring semantic language reasoning from deterministic computation and information-retrieval infrastructure. Model selection and evaluation must then confirm that the chosen Claude model satisfies quality, latency, and cost requirements.

Yes, Yes, Yes, No, No Dense vector retrieval represents semantic similarity, making it suitable for matching paraphrases and conceptually related language even when the query and source do not share identical words. Sparse lexical retrieval, including BM25-style matching, retains strong sensitivity to exact tokens and is therefore valuable for identifiers, product codes, technical names, and uncommon terminology. Structured query retrieval is appropriate when the required operation depends on explicit fields, predicates, joins, counts, grouping, or aggregation over a relational schem
a. In that situation, generating or invoking a constrained database query is more precise than approximating the operation through semantic similarity. Hybrid retrieval is not identical to dense retrieval. It combines semantic and lexical candidate sets, typically followed by rank fusion or reranking. Anthropic’s contextual-retrieval guidance explains that semantic search captures meaning and paraphrases, while BM25 captures exact terminology; combining them improves coverage. Contextual Retrieval, Contextual Retrieval Cookbook Random sampling is not a relevance strategy. It provides no systematic relationship between the query and selected evidence, producing unstable coverage and preventable hallucination risk. Production Q&A requires deterministic or evaluated relevance mechanisms, access filters, suitable indexes, and measurable retrieval metrics such as recall at k. Study Guide

/topics: Dense retrieval; sparse retrieval; structured queries; hybrid search;

rank fusion; retrieval evaluation; production RAG design.

===============

Yes, Yes, No, Yes, No Zero-shot prompting is a reasonable baseline for a clearly defined closed-set classifier because the model receives both the permissible labels and their meanings. Performance should then be measured against an evaluation set before additional prompt complexity is introduced. Few-shot examples are also sound for extraction because examples containing the exact field names demonstrate the required schema and reduce formatting ambiguity. Anthropic identifies examples as one of the most reliable methods for steering output format, structure, and consistency. Prompting Best Practices Chain-of-thought is not an unconditional default. Additional reasoning is appropriate when the task requires substantive decomposition or complex judgment, but routine classification and extraction may gain little while consuming more tokens and increasing latency. The technique should be selected according to evaluated task requirements. An even-handed comparison governed by explicit criteria mitigates leading-question bias by requiring symmetrical treatment of alternatives rather than presupposing a preferred conclusion. Finally, prompt validation is model-specific. New model versions may follow instructions differently, change formatting behavior, alter tool selection, or produce different latency and cost characteristics. Anthropic’s migration guidance consequently recommends reevaluating prompts and workloads against the new baseline. Model Migration Guide Study Guide

/topics: Zero-shot prompting; few-shot examples; reasoning techniques;

neutral framing; model migration and regression evaluation.

===============

When a RAG (Retrieval-Augmented Generation) application suddenly begins hallucinating or contradicting sources without any changes to the retrieval pipeline, the root cause typically lies in the generation phase. The two primary variables are the model version and the system prompt. Testing the previous model version (C) isolates whether a recent model update or deprecation caused a regression in instruction-following capabilities. Inspecting the system prompt (E) verifies whether grounding instructions or citation constraints were accidentally altered, which directly controls the model's adherence to the provided context.

A. Increasing the context window does not diagnose the root cause and may exacerbate hallucinations by introducing more noise into the prompt.
B. The scenario explicitly states there was no change to the retrieval pipeline, making index changes irrelevant to diagnosing the sudden degradation.
D. Reducing temperature mitigates variance but acts as a band-aid rather than diagnosing the underlying root cause of the sudden behavioral shift.

1. Anthropic. (2024). Claude API Documentation: Prompt Engineering for RAG. Section: Grounding and Citation Constraints, Paragraph 2.

2. Anthropic. (2024). Claude API Documentation: Model Upgrades and Versioning. Section: Regression Testing and Version Pinning.

3. Lewis, P., et al. (2020). "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." Advances in Neural Information Processing Systems, 33, 9459-9474. https://doi.org/10.48550/arXiv.2005.11401

Authentication and authorization run before retrieval so retrieval can filter by identity — Yes An asynchronous queue absorbs bursty traffic between intake and the model-invocation layer — Yes Tax computation is encoded directly in the system prompt rather than in code — No Conversation logs include unredacted government identifiers to maximize tuning signal — No Authentication and authorization must precede retrieval so unauthorized information never enters model context. An asynchronous queue is appropriate for burst absorption, backpressure, controlled concurrency, and retry management at scale. Deterministic tax computation should be implemented in validated code or a controlled calculation tool, not delegated to probabilistic prompt interpretation. Including unredacted government identifiers in conversation logs violates data- minimization principles and creates unnecessary privacy, security, and regulatory exposure. Sensitive fields should be removed, tokenized, or redacted before logging or model use unless specifically required and authorized. The resulting design separates deterministic computation, access control, scalable orchestration, and language reasoning into appropriate architectural layers.