Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 29

Question 280

A team upgrades models while keeping the same named effort setting. What should it avoid assuming?

  1. Identical behavior from the same setting across releases.
  2. That settings should be recorded.
  3. That supported combinations should be verified.
  4. That the task’s acceptance criteria still apply.

Explanation

Correct Answer

A:

The same setting label can map to different behavior on another release, so retain configuration records and run workload-specific comparisons.

Incorrect Answers

B:

Recording them remains useful for comparison.

C:

Compatibility checks remain necessary during migration.

D:

Business requirements continue to govern the upgraded system.

Question 281

A supported fast mode reduces latency but costs more. When is enabling it justified?

  1. When the workload has no latency requirement and budget is already exceeded.
  2. When measured latency benefits meet the product need within its cost constraints.
  3. Only because higher prices guarantee factual correctness.
  4. Whenever its name includes fast.

Explanation

Correct Answer

B:

A faster mode is a tradeoff, not an automatic optimization for every workload. Compare the measured reduction in latency with its additional cost and the product's performance requirement instead of assuming faster is always economical.

Incorrect Answers

A:

Those constraints provide no demonstrated justification for the additional expense.

C:

Price is not proof of accuracy on the application’s tasks.

D:

A label does not establish an acceptable cost-benefit relationship.

Question 282

Human reviewers consistently favor answers labeled with a prestigious model name, even when identical answers receive lower scores under another label. What evaluation change addresses this bias?

  1. Show larger model names in bold to help reviewers choose.
  2. Use each model's marketing claims as its quality score.
  3. Blind model identities and randomize presentation order during review.
  4. Discard ratings only when the favored model loses.

Explanation

Correct Answer

C:

Removing brand cues and balancing order reduces evaluation effects unrelated to the actual answer quality being compared.

Incorrect Answers

A:

Emphasizing the cue strengthens the observed bias instead of isolating answer quality.

B:

Promotional descriptions do not measure performance on the team's evaluated responses.

D:

Selective exclusion biases the results further and invalidates the comparison.

Question 283

Increasing model size does not improve a task because retrieval supplies the wrong documents. What should be fixed first?

  1. Always the output formatting.
  2. The evidence-selection problem in retrieval.
  3. Only the sampling temperature while retaining the same incorrect evidence.
  4. Only the response parser, after confirming it faithfully preserves model output.

Explanation

Correct Answer

B:

A stronger model cannot reliably answer from the wrong evidence; repair retrieval relevance before attributing the failure to model capacity.

Incorrect Answers

A:

Formatting does not repair the demonstrated evidence defect.

C:

Sampling changes cannot reliably compensate for retrieving the wrong source documents.

D:

A correct parser cannot repair irrelevant documents supplied before generation; the observed constraint lies in retrieval.

Question 284

A small model answers easy cases and escalates others. Which metric is needed beyond its direct-answer accuracy?

  1. Assume escalation removes every error automatically.
  2. Only the shortest generated answer.
  3. The routing/escalation coverage and errors on cases it incorrectly keeps.
  4. Only the smaller model's cost on requests it answers directly.

Explanation

Correct Answer

C:

Measure coverage, mistaken non-escalations, and final outcomes as well as direct-answer accuracy. A router may look accurate simply by forwarding almost everything, sacrificing the intended efficiency.

Incorrect Answers

A:

The router can fail to escalate cases that need stronger handling.

B:

Response length does not show whether escalation decisions are correct.

D:

That omits escalation frequency and expensive second-stage work, understating the router's full operational cost.

Question 285

Under token-based billing, a request allows 4,000 output tokens but generates 300. Which quantity should cost estimation use?

  1. Estimate output tokens as exactly the number of visible characters.
  2. Actual reported usage under the applicable pricing rules.
  3. Always the entire 4,000-token allowance.
  4. Zero because the response ended early.

Explanation

Correct Answer

B:

Use actual reported billable usage and applicable rates. The configured output ceiling is a capacity allowance, not evidence that the entire allowance was generated.

Incorrect Answers

A:

Tokenization is not a one-character-per-token mapping, and billable categories should come from reported usage.

C:

Unused capacity is not equivalent to observed generation.

D:

A short response still consumes resources.

Question 286

A hypothetical cache setup costs 8 units and saves 2 units on each subsequent reuse. When do cumulative savings exceed setup cost?

  1. After five reuses.
  2. Before any reuse.
  3. After four reuses.
  4. After one reuse.

Explanation

Correct Answer

A:

Four reuses save 8 units and exactly recover setup cost. Five reuses save 10 units, the first point at which savings exceed that cost.

Incorrect Answers

B:

No reuse savings have occurred yet to offset the setup cost.

C:

That is the break-even point, not a strictly positive net saving.

D:

Two saved units are less than the eight-unit setup cost.

Question 287

Model A is cheaper per call but frequently requires retries and manual correction. What is a better comparison metric?

  1. Only the price per initial model call, excluding retries.
  2. Only the advertised per-call price.
  3. Total cost per acceptable completed task.
  4. Only the final answer’s length.

Explanation

Correct Answer

C:

Total task economics include recovery and correction rather than only the first inference charge. Include retries, fallback calls, and correction effort when comparing the resources required to deliver one acceptable completed result.

Incorrect Answers

A:

The cheaper first attempt can still lead to greater total expense when accepted outcomes require repeated inference or human correction.

B:

A lower initial price can be outweighed by repeated calls and operational effort.

D:

Length does not capture retries, failures, or human work.

Question 288

Request lengths vary greatly. How should a monthly cost forecast represent this uncertainty?

  1. Assume every request uses the maximum possible context.
  2. Ignore output usage entirely.
  3. Use observed workload distributions and scenario ranges.
  4. Multiply one unusually short request by all traffic.

Explanation

Correct Answer

C:

A realistic forecast accounts for varying input, output, and traffic rather than assuming every request matches one convenient sample.

Incorrect Answers

A:

This may provide an extreme bound but is not a representative expected forecast.

B:

Generated tokens can materially contribute to total cost.

D:

That can substantially underestimate the actual workload.

Question 289

A caching change is expected to save money. Which evidence should verify that expectation?

  1. Only the presence of cache_control in source code.
  2. Only the prompt’s total length.
  3. Reported cache reads/writes and end-to-end cost under representative reuse.
  4. Only one cold request.

Explanation

Correct Answer

C:

Cache eligibility alone does not establish actual hits or net savings after write overhead. Inspect actual cache-read and cache-write usage alongside total spend; configuring a cache breakpoint alone does not prove requests reused it.

Incorrect Answers

A:

Configuration intent does not prove runtime reuse.

B:

Length affects eligibility and cost but does not measure achieved hits.

D:

A cold request cannot demonstrate the benefits of repeated reuse.

Question 290

A system hits its budget cap even though request rate is low. What distinction matters?

  1. A low request count guarantees low token cost.
  2. Treat the budget failure as a network timeout and retry indefinitely.
  3. Every limit is solved by millisecond jitter.
  4. Cumulative spending limits differ from instantaneous throughput limits.

Explanation

Correct Answer

D:

Low traffic does not undo accumulated spend; recovery must address the actual budget condition. A few large requests can consume a financial allowance without approaching a requests-per-minute limit, so diagnose which quota was actually reached.

Incorrect Answers

A:

Large or expensive requests can consume substantial budget.

B:

A financial allowance is not restored by network retry, and repeated attempts can worsen cost or remain consistently blocked.

C:

Delay variation cannot reset an exhausted spending allowance.

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