Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 29

Question 280

A codebase contains tens of thousands of files, and Claude Code exhausts context by reading entire directories. What investigation pattern is more effective?

  1. Search narrowly, then read only relevant symbols
  2. Skip source inspection and rely on general framework familiarity
  3. Paste the entire repository into every prompt
  4. Modify files before learning which modules are involved

Explanation

Correct Answer

A:

Selective retrieval keeps the active context focused. Directory maps and file references help explore deeply only where evidence warrants.

Incorrect Answers

B:

Repository-specific behavior cannot be inferred reliably from generic knowledge. Actual code matters.

C:

Unfiltered code consumes context and can bury the relevant paths. It also creates avoidable cost.

D:

Premature edits risk breaking dependent components. Orientation should precede mutation.

Question 281

A developer wants the import chain for one public endpoint in a very large monorepo. Which exploration step should come first?

  1. Ask Claude to invent the likely call graph
  2. Locate the endpoint definition and follow only its referenced modules
  3. Read every module alphabetically before searching for the endpoint
  4. Perform a repo-wide refactor while mapping imports

Explanation

Correct Answer

B:

A dependency-driven search narrows context to the relevant call graph. The investigator can expand only where a link requires it.

Incorrect Answers

A:

A guess cannot establish actual dependencies in the current repository. Source references must be inspected.

C:

Alphabetical full scans waste time and context. The entry point already provides a precise target.

D:

Read-only mapping does not authorize changes. Mutation would introduce noise into the investigation.

Question 282

A coding agent frequently forgets earlier discovered file paths during a prolonged refactor. What artifact should it maintain?

  1. A list of filenames chosen before inspecting the code
  2. A claim that the entire repository has been understood
  3. A persistent map of files, dependencies and pending edits
  4. Only an unstructured copy of all previous tool outputs

Explanation

Correct Answer

C:

A project map survives context compression and helps resume accurately. It should link to source rather than reproduce every file.

Incorrect Answers

A:

Unverified guesses can misdirect subsequent steps. The map should be updated from actual exploration.

B:

Unsupported certainty does not preserve practical navigation. Precise dependencies and gaps matter.

D:

Raw transcripts are expensive and hard to search. A curated map better supports continuation.

Question 283

A team delegates code search to subagents but the coordinator is overwhelmed by verbose outputs. What return contract should be adopted?

  1. Unattributed code snippets with no repository location
  2. Full raw search logs from every worker on every turn
  3. Return concise findings with paths, evidence and blockers
  4. Only yes or no without any locations

Explanation

Correct Answer

C:

The parent needs inspectable pointers rather than every search step. Scoped handoffs preserve context for synthesis and implementation.

Incorrect Answers

A:

Contextless snippets cannot be checked reliably against current files. Provenance is part of the handoff.

B:

Verbose logs duplicate information and consume the coordinator's context. They are better stored outside the active prompt.

D:

Binary answers are insufficient for review and editing. Evidence paths make results actionable.

Question 284

A developer needs to understand how authentication flows through an unfamiliar project. What source-first strategy is best?

  1. Trace entry points, middleware and identity checks
  2. Read only files with auth in their names
  3. Assume every route inherits the same security behavior
  4. Disable tests to make exploration less expensive

Explanation

Correct Answer

A:

Following the actual execution path reveals where authentication and authorization occur. It avoids assumptions based on filenames alone.

Incorrect Answers

B:

Critical checks may reside in middleware or shared components with different names. A call-chain approach is broader and better grounded.

C:

Routing and middleware exceptions can differ. Actual code should establish each boundary.

D:

Removing test coverage does not help understand the flow and weakens later verification. Investigation should remain read-only.

Question 285

A patch spans several packages, and a context compaction drops the original acceptance tests. What should be preserved outside chat?

  1. Only the most recent changed file name
  2. A guess that all tests passed before context was lost
  3. A package-level task checklist tied to required tests
  4. Every redundant debug line from previous sessions

Explanation

Correct Answer

C:

A durable progress artifact prevents loss of acceptance criteria. The agent can reconcile completed work after compaction or restart.

Incorrect Answers

A:

A multi-package task needs cross-component requirements. One filename omits dependencies and verification.

B:

Missing evidence is not proof of success. Test state must be recorded from actual runs.

D:

Bulk transcripts are less useful than a focused acceptance ledger. Preserve what controls the next decision.

Question 286

A code exploration agent needs a summary of the repository's architecture, but not the contents of every file. What result format is suitable?

  1. A raw concatenation of the full repository
  2. A component map with entry points, boundaries and citations
  3. An unsupported list of modules invented for completeness
  4. A generic diagram based on the programming language alone

Explanation

Correct Answer

B:

High-level structure with file anchors is useful for planning and can be verified selectively. It balances coverage against context budget.

Incorrect Answers

A:

A giant dump is not an architecture summary and may exceed the context window. It obscures important relationships.

C:

Fabricated components can mislead refactors. Unknown areas should be marked as such.

D:

Language choice does not determine a project's actual architecture. The summary must reflect inspected code.

Question 287

An AI reviewer recommends deleting production customer records but says it is 95 percent confident. What control is most appropriate?

  1. Ask the model to raise its stated confidence to 100 percent
  2. Execute immediately because 95 percent sounds high
  3. Hide the proposed deletion from the change-review log
  4. Human approval and verification before irreversible deletion

Explanation

Correct Answer

D:

High-impact destructive actions deserve accountable approval. Model confidence is not a substitute for permissions, evidence or impact assessment.

Incorrect Answers

A:

Changing a label does not change the underlying uncertainty. The system needs external verification.

B:

A numerical self-report does not certify safety. Even a small error risk is consequential for irreversible data loss.

C:

Concealment undermines accountability and makes recovery harder. The decision must be auditable.

Question 288

A document extractor reports high confidence for some invoices but consistently fails on handwritten totals. How should confidence calibration be evaluated?

  1. Set every response confidence to high for consistency
  2. Check confidence against labeled handwritten examples
  3. Trust model confidence without measuring accuracy
  4. Remove difficult handwritten examples from tests

Explanation

Correct Answer

B:

Empirical error rates reveal where confidence is misleading. Domain-specific evaluation can inform review thresholds.

Incorrect Answers

A:

Uniform labels contain no useful risk information. Calibration should reflect observed reliability.

C:

Self-reported certainty may be poorly calibrated on unusual input. A validation dataset is required.

D:

Excluding the failure class hides the true production risk. The test set should represent realistic inputs.

Question 289

A team wants a human to review only invoices likely to contain material extraction errors. What routing design is suitable?

  1. Review only the longest invoices regardless of field quality
  2. Review none because the JSON schema passes
  3. Let the model secretly decide its own exceptions without audit
  4. Use risk thresholds, targeted review and periodic sampling

Explanation

Correct Answer

D:

Review effort can focus on uncertain or high-impact records while samples monitor errors among auto-approved cases. Thresholds should come from measured performance.

Incorrect Answers

A:

Length may correlate with difficulty but is not a complete risk signal. Evaluate evidence and business impact.

B:

Shape checks cannot detect incorrect totals or missing source evidence. Semantic risk remains.

C:

Unobservable routing cannot be calibrated or challenged. The decision criteria should be documented and measurable.

Question 290

A customer service assistant must always escalate potential fraud accusations before taking adverse action. Which mechanism is best?

  1. Hide the source evidence from the reviewer to avoid bias
  2. A mandatory review gate triggered by explicit risk criteria
  3. Allow the assistant to bypass review for familiar customers
  4. Treat the model's fraud label as an automatic ban

Explanation

Correct Answer

B:

A human checkpoint can assess evidence and consequences before account action. The agent can gather facts but not make irreversible accusations alone.

Incorrect Answers

A:

Reviewers need relevant facts to assess the claim. Data should be scoped and presented with provenance.

C:

Familiarity is not a documented exception and can create inconsistent treatment. The gate should apply according to policy.

D:

An unverified classifier can produce false positives. Human review is the stated requirement.

Claude Certified Architect - Foundations (CCAR-F) (CCA-F) Practice Tests