CCAR-F 시험 - Anthropic실제시험문제와 답 - 320문항

Question No : 1

Which of the following are recommended architectural practices for preserving information provenance and handling uncertainty in a multi-agent system at quicktechie.com? Choose 2 correct answers.

A.Requiring subagents to output structured claim-source mappings (including source URLs, document names, and excerpts) that downstream agents must preserve.
B.Silently dropping contested findings to ensure the final synthesis report appears highly confident.
C.Structuring reports with explicit sections that distinguish well-established findings from contested ones, preserving the original methodological context.
D.Omitting publication dates from the metadata to save valuable token space in the coordinator's context window.

정답:
Explanation:
Recommended practices for information provenance include requiring structured claim-source mappings that pass through to the final output, and structuring reports to explicitly distinguish well-established findings from contested ones. Dropping contested findings or omitting temporal metadata are anti-patterns that destroy critical context.

Question No : 2

When designing the output structure for a synthesis agent that processes diverse data streams (e.g., financial statistics, news prose, and technical specifications), what is the recommended presentation pattern?

A.Convert all extracted data into a uniform, flat JSON array to simplify downstream parsing.
B.Render different content types appropriately (e.g., financial data as tables, news as prose, technical findings as structured lists) rather than forcing a uniform format.
C.Use the Code Execution tool to automatically generate an Excel workbook containing all data types.
D.Restrict the output strictly to natural language paragraphs to maintain conversational fluency.

정답:
Explanation:
To preserve the utility and readability of synthesized multi-source data, different content types should be rendered appropriately (e.g., financial data as tables, news as prose, technical findings as structured lists) rather than destructively converting everything into a single uniform format.

Question No : 3

During a multi-source research workflow, what is the primary risk to information provenance when intermediate findings are compressed using standard summarization?

A.The model's extended thinking budget is exhausted before synthesis can occur.
B.Source attribution is silently lost unless structured claim-source mappings are explicitly required and preserved.
C.The context window automatically truncates the middle sections of the summary.
D.It triggers a PreToolUse hook that blocks the synthesis agent from reading the data.

정답:
Explanation:
A major risk in multi-source synthesis is that source attribution is lost during summarization steps when findings are compressed. To prevent this, architects must require structured claim-source mappings that downstream agents are instructed to explicitly preserve through the synthesis phase.

Question No : 4

Why is it critical to require subagents to include publication or data collection dates in their structured outputs before passing them to a downstream synthesis agent?

A.To automatically trigger the Message Batches API's 24-hour processing window logic.
B.To prevent temporal differences in the underlying data from being misinterpreted by the synthesis agent as direct factual contradictions.
C.To ensure strict ISO 8601 validation passes when using the tool_choice: " any" parameter.
D.To sort the JSON keys alphabetically and reduce context window token consumption.

정답:
Explanation:
Requiring publication or collection dates in structured outputs is essential because it prevents temporal differences (e.g., a 2022 revenue report vs a 2024 revenue report) from being misinterpreted by the synthesis agent as a contradiction in facts.

Question No : 5

When two subagents at quicktechie.com provide conflicting statistical data from different credible sources during a synthesis task, what is the architecturally correct pattern for handling this uncertainty?

A.Arbitrarily select the value from the most recently updated source to resolve the conflict smoothly.
B.Silently average the conflicting numerical values to provide a balanced estimate to the user.
C.Preserve both values and explicitly annotate the conflict with source attribution, letting the coordinator or user reconcile it.
D.Drop the contested data point entirely to ensure the final report only contains unified facts.

정답:
Explanation:
When handling conflicting statistics from credible sources, the correct architectural pattern is to annotate the conflicts with explicit source attribution and methodological context rather than arbitrarily selecting one value or silently averaging them. This preserves information provenance for informed downstream resolution.

Question No : 6

When designing a human review workflow with limited reviewer capacity, which of the following items should be prioritized for human review? (Choose 2 correct answers)

A.Extractions where the model reports low field-level confidence.
B.Randomly selected extractions from documents categorized as having 99% historical accuracy.
C.Extractions involving ambiguous or contradictory source documents.
D.Documents where the tool_choice was set to "auto".

정답:
Explanation:
When reviewer capacity is limited, human review should be prioritized and routed toward items with low model confidence and tasks involving ambiguous, contradictory, or borderline source material where human judgment is strictly required.

Question No : 7

Before automating the extraction of high-confidence fields in a quicktechie.com processing system, what validation step is crucial?

A.Validating the JSON schema syntax using a PreToolUse hook.
B.Validating accuracy segmented by document type and specific field to verify consistent performance across all segments.
C.Confirming consensus through 3 separate monolithic review passes.
D.Checking that the model's self-reported sentiment analysis is positive.

정답:
Explanation:
Before reducing human review and automating high-confidence extractions, it is crucial to validate accuracy segmented by both document type and specific field. This ensures that performance is actually consistent across all segments, rather than relying on misleading aggregate averages.

Question No : 8

To measure ongoing error rates in high-confidence extractions and detect novel error patterns, which methodology should be implemented?

A.Stratified random sampling.
B.Exhaustive 100% human review.
C.Aggregate accuracy averaging.
D.Prompt-based sequential pipelining.

정답:
Explanation:
Stratified random sampling is the recommended method for measuring error rates in data classified as 'high-confidence'. This catches novel error patterns that the model might be confidently hallucinating without requiring unscalable 100% manual review.

Question No : 9

How should field-level confidence scores from a model be utilized to optimize a human-in-the-loop workflow?

A.By trusting the model's self-reported conversational confidence score to bypass human review.
B.By calibrating the scores using labeled validation sets to effectively route review attention to low-confidence extractions.
C.By using them to automatically increase the max_tokens limit on subsequent runs.
D.By setting tool_choice: 'any" whenever confidence drops below 0.5.

정답:
Explanation:
To optimize human review, models should output structured field-level confidence scores that are rigorously calibrated against labeled validation sets. This ensures review thresholds are accurate, allowing high-risk or low-confidence findings to be correctly routed to human

Question No : 10

An automated data extraction pipeline reports 96% aggregate accuracy, but downstream users complain about frequent errors on unstructured contracts.
What is the most likely cause of this discrepancy?

A.The system is experiencing the 'lost in the middle' effect.
B.Aggregate accuracy metrics are masking poor performance on specific document types (e.g., contracts vs receipts).
C.The model's temperature parameter is set too low for contract analysis.
D.The pipeline lacks a PostToolUse hook for error catching.

정답:
Explanation:
Aggregate accuracy metrics can mask severe per-document-type failures. If standard receipts perform at 99% accuracy but complex contracts perform at 70%, the aggregate may still look high. Teams must track accuracy by document type (stratified metrics) to identify hidden failures.

Question No : 11

Which of the following represent best practices for managing context during a large codebase exploration? (Choose 2 correct answers)

A.Relying solely on the model's extended thinking capabilities instead of external scratchpad files.
B.Summarizing key findings from one exploration phase and injecting those summaries into the initial context for the next phase's subagents.
C.Passing the full, uncompressed conversation history of the coordinator to every subagent.
D.Using subagent delegation to isolate verbose search results and file reads from the coordinator's context.

정답:
Explanation:
Best practices include delegating verbose tasks to subagents to preserve coordinator context, and injecting summarized findings from previous phases into the context of newly spawned subagents. Passing full conversation history causes context pollution, and abandoning scratchpad files risks permanent context degradation.

Question No : 12

In an extended exploration session, which built-in Claude Code command is explicitly used to reduce context usage when the window fills with verbose discovery output?

A./clear
B./compact
C.- - resume
D./init

정답:
Explanation:
The /compact command is specifically designed to compress conversation history and reclaim space in the context window while preserving essential knowledge. /clear completely wipes the context, which is not desirable if you want to retain project knowledge.

Question No : 13

How should an architect design crash recovery for a multi-agent exploration system at quicktechie.com?

A.Use structured agent state exports (manifests) that the coordinator loads on resume and injects into agent prompts.
B.Rely entirely on the 4. claude .json file to store conversational memory.
C.Parse natural language checkpoints directly from the assistant's unstructured text responses.
D.Automatically trigger a /clear command to gracefully reset the environment.

정답:
Explanation:
Designing crash recovery requires persistent, structured state. Best practice dictates that each agent exports its state to a known location as a manifest, which the coordinator can then load and inject upon resumption to seamlessly recover the session state.

Question No : 14

To keep the main agent's context clean while investigating a complex authentication flow, which architectural pattern is highly recommended?

A.Spawn a subagent to isolate the verbose exploration output, returning only a focused summary to the main agent.
B.Use the Message Batches API to process the exploration synchronously.
C.Apply progressive summarization to the main conversation history every 2 turns.
D.Use the fork_session command to duplicate the main agent permanently.

정답:
Explanation:
Subagent delegation is the recommended pattern to isolate verbose exploration output (like reading multiple files and tracing functions) from the main coordinator. The subagent does the heavy lifting in its own isolated context and returns only a clean synthesis to the main agent.

Question No : 15

During an extended codebase exploration session in Claude Code, context degradation begins to occur.
What is the most effective strategy to persist key findings across context boundaries?

A.Use the - -json-schema flag to aggressively compress output.
B.Have agents maintain scratchpad files to record key findings and reference them in subsequent questions.
C.Increase the model's temperature parameter to recall lost context.
D.Continuously re-feed the entire codebase into the prompt every 5 turns.

정답:
Explanation:
Context degradation in extended sessions causes models to lose track of earlier details. Using scratchpad files to persist critical intermediate state externally ensures that key findings survive context compression and session boundaries.

Anthropic CCAR-F 시험