Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 14

Question 130

A scanned PDF contains text only as page images. Plain text extraction returns empty strings. What should ingestion use?

  1. A supported visual-document path or validated OCR extraction.
  2. Treat empty extracted text as proof the document is blank.
  3. Increase the language model’s temperature.
  4. Pass only the PDF’s file extension.

Explanation

Correct Answer

A:

The information exists visually rather than in a text layer. Ingestion must access that representation before downstream reasoning.

Incorrect Answers

B:

A missing text layer does not mean the rendered pages contain no information.

C:

Sampling changes cannot supply document content that ingestion never provided.

D:

The extension identifies format but contains none of the page evidence.

Question 131

A direct Anthropic API integration is moved to a cloud-provider endpoint. What must be checked first?

  1. The endpoint’s required authentication and model identifier conventions.
  2. Whether the model family appears in both marketing pages.
  3. Whether the same prompt text looks attractive.
  4. Whether the old API key has enough characters.

Explanation

Correct Answer

A:

Provider integrations can use different credentials and resource naming. Direct-API assumptions cannot be copied without verification.

Incorrect Answers

B:

Shared branding does not prove identical request or authorization contracts.

C:

Prompt presentation does not establish endpoint authentication compatibility.

D:

Credential validity depends on the endpoint’s authentication scheme, not string length alone.

Question 132

A feature works on one Claude access platform. What justifies enabling it on another?

  1. Current documentation and a compatibility test for that platform and model.
  2. The same feature parameter is accepted without inspecting the returned behavior.
  3. The same model-family label appears in both platforms' catalogs.
  4. Both endpoints accept ordinary Messages-style text requests.

Explanation

Correct Answer

A:

Availability can differ across surfaces. Verify the specific integration rather than assuming all Claude endpoints expose identical features.

Incorrect Answers

B:

Acceptance alone may not establish equivalent semantics; confirm documented support and test the capability on the target platform.

C:

A shared family label does not establish identical feature exposure, rollout, or request semantics on two access platforms.

D:

Basic text generation compatibility does not establish support for the particular additional feature being migrated.

Question 133

A supported thinking configuration consumes output budget beyond visible answer text. What should usage reporting use?

  1. Only words shown in the final answer.
  2. Estimate usage only from the final answer's visible word count.
  3. A fixed charge of zero for invisible output.
  4. The API’s reported usage rather than displayed character count.

Explanation

Correct Answer

D:

Visible text can omit billable generation categories. Reported token usage reflects the provider’s accounting more appropriately than UI length.

Incorrect Answers

A:

That can miss generated tokens not displayed as ordinary answer prose.

B:

Visible words omit reasoning-related token usage and do not map exactly to billable token categories.

C:

Visibility does not determine whether a generation category is billed.

Question 134

A downstream service requires machine-readable JSON matching a supported schema. Which API capability should be evaluated?

  1. Rendering JSON-looking text as an image.
  2. Native structured outputs with explicit validation in the consumer.
  3. Removing the schema from the request.
  4. Only a request to “sound like JSON.”

Explanation

Correct Answer

B:

A constrained output format addresses structural reliability; the consumer still handles completion conditions and business correctness.

Incorrect Answers

A:

An image does not satisfy the downstream machine-readable JSON interface.

C:

Without the contract, the model has less information about the required output structure.

D:

Stylistic guidance is weaker than a supported structured-output contract.

Question 135

An application wants traceable claims from supplied documents. What should its request design preserve?

  1. Assign citation numbers after generation without linking them to supplied passages.
  2. Only the document titles without contents.
  3. Document identity and supported citation-capable source structure.
  4. Only a combined anonymous text blob.

Explanation

Correct Answer

C:

Traceability requires a link between claims and the specific supplied evidence, not merely plausible-looking footnote numbers.

Incorrect Answers

A:

Post-hoc numbering can label claims but cannot establish which actual source supports each one.

B:

Titles alone usually do not supply the facts needed to substantiate detailed claims.

D:

Removing source boundaries makes precise attribution harder.

Question 136

An application uploads a document once and reuses its file reference in many supported Messages requests. The team assumes every later request processes zero document input tokens. What is wrong with that assumption?

  1. File reuse and prompt-cache accounting are separate mechanisms.
  2. Only the file identifier string can contribute to input-token usage.
  3. Uploading a file automatically guarantees a permanent prompt-cache hit.
  4. Uploading once makes the document eligible for unlimited free inference.

Explanation

Correct Answer

A:

A reusable reference avoids repeating the upload, but the document can still be supplied as model context. Input usage and any eligible cache treatment must be checked separately rather than inferred from the short identifier.

Incorrect Answers

B:

The referenced content is provided to Messages for processing, not merely the identifier characters. File reuse does not make the document disappear from input accounting.

C:

Storage availability does not establish cache eligibility, matching prefixes, or cache lifetime. Those conditions belong to a distinct feature.

D:

Free or reusable storage operations are distinct from processing the document in model requests. Applicable input and cache accounting still matter.

Question 137

Why can adding many unused tool definitions increase request cost even when Claude invokes none?

  1. Unused tools are always charged as successful external transactions.
  2. Every declared tool necessarily executes once.
  3. The server automatically reserves one tool execution for each definition.
  4. Definitions contribute to the model’s input context.

Explanation

Correct Answer

D:

Tool descriptions and schemas enter model input processing before any invocation. Account for this input-token overhead separately from execution-specific charges.

Incorrect Answers

A:

External execution charges require the relevant execution; input-token overhead is a separate mechanism.

B:

Declaration does not imply execution; the cost increase can arise from input alone.

C:

A supplied definition does not imply an execution reservation. Input processing can account for the increase even without calls.

Question 138

A supported server-tool response returns pause_turn. What should an application distinguish it from?

  1. An instruction to fabricate all missing tool results.
  2. A final completed user answer; follow the documented continuation flow.
  3. A guarantee that the request changed no external state.
  4. Proof that the API credential is invalid.

Explanation

Correct Answer

B:

A paused server-tool turn can require another request to continue work rather than treating the partial response as final.

Incorrect Answers

A:

Continuation should preserve actual evidence, not invent observations for unfinished server work.

C:

Pause status describes control flow, not a universal rollback or side-effect guarantee.

D:

A generation stop condition is different from an authentication failure.

Question 139

An async model-call function is invoked without await, and its return value is treated as the completed response. What is wrong?

  1. Calling the coroutine function already starts and completes the request.
  2. Access response attributes directly on the returned coroutine object.
  3. Sleeping briefly makes the unscheduled coroutine become a response.
  4. The coroutine must be awaited or scheduled and its completion observed.

Explanation

Correct Answer

D:

Calling an async function returns a coroutine object. Execute it through the event loop and observe its completed value before accessing response fields.

Incorrect Answers

A:

Calling an async function produces a coroutine object. Its execution and eventual result still depend on scheduling and observing completion.

B:

A coroutine is not the completed API result; it must be scheduled and awaited before its response fields exist.

C:

Elapsed time alone does not schedule the coroutine or obtain its return value. The event loop must execute it and the caller must observe completion.

Question 140

A large local document conversion blocks the Python event loop before any API call. What should change?

  1. Increase model output tokens.
  2. Move blocking computation to an appropriate worker or execution facility.
  3. Increase the number of asynchronous tasks on the same blocked event loop.
  4. Add async to the function name only.

Explanation

Correct Answer

B:

Move substantial CPU work to a suitable worker or process so other scheduled requests can make progress while conversion runs.

Incorrect Answers

A:

The stall occurs locally before inference, so generation limits are irrelevant.

C:

Scheduling more tasks cannot make progress while synchronous CPU work monopolizes that loop's execution thread.

D:

Naming does not change blocking computation into cooperative asynchronous work.

Claude Certified Developer - Foundations (CCDV-F) Practice Tests