Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 25

Question 240

A supported sampling configuration is adjusted to encourage more varied wording. What tradeoff should be evaluated?

  1. Whether increased diversity affects consistency on the task.
  2. Whether vocabulary diversity alone is sufficient to prove better task performance.
  3. Whether a single attractive sample proves all outputs satisfy the quality target.
  4. Whether every possible response becomes equally likely regardless of the prompt.

Explanation

Correct Answer

A:

Evaluate diversity against the task's consistency requirements. Variation useful for creative exploration can reduce repeatability in constrained classification or extraction.

Incorrect Answers

B:

More varied wording can coexist with worse accuracy or format compliance, so diversity is not a complete quality measure.

C:

One selected result cannot characterize the broader output distribution or detect failures introduced by increased variation.

D:

Sampling controls alter selection probabilities; they do not erase input conditioning or make the entire response space uniform.

Question 241

For this simplified sampler, tokens have probabilities 0.50, 0.30, 0.15, and 0.05. It retains the smallest descending-probability set whose cumulative mass reaches at least 0.70. Which set remains?

  1. Only the token with probability 0.50.
  2. All four tokens because their probabilities sum to one.
  3. The tokens with probabilities 0.50 and 0.30.
  4. The tokens with probabilities 0.15 and 0.05.

Explanation

Correct Answer

C:

The first token alone contributes 0.50, below the threshold. Adding the second brings cumulative probability to 0.80, so the specified smallest qualifying set contains two tokens.

Incorrect Answers

A:

The most likely token is not enough under the stated cumulative threshold. Its probability mass remains 0.20 below the requirement.

B:

The rule asks for the smallest qualifying set, not the complete distribution. Two tokens already reach the threshold.

D:

Their combined mass is only 0.20, and they are not the leading tokens in descending-probability order.

Question 242

A generative model produces a highly probable continuation containing a specific historical date. What does the continuation's high token probability establish?

  1. The date has been checked against an authoritative live database.
  2. It is favored by the model's conditional distribution, not independently verified.
  3. Every alternative date is impossible in the real world.
  4. The model is reporting a calibrated probability that the date is true.

Explanation

Correct Answer

B:

Generation likelihood reflects the model's learned patterns and current context. A plausible factual sequence still needs appropriate evidence when exact historical accuracy matters.

Incorrect Answers

A:

Sampling probability does not establish that retrieval or a database lookup occurred. External verification requires an actual evidence source.

C:

A preferred continuation does not rule out factual alternatives. Likelihood ranking and the truth of historical events are different questions.

D:

Token likelihood measures a continuation under the model, not automatically a calibrated factual-correctness score for the resulting claim.

Question 243

A supported model exposes effort settings. How should a team choose one for a workload?

  1. Always choose maximum without measurement.
  2. Assume effort has no operational effect.
  3. Compare quality, latency, and usage across supported settings on representative tasks.
  4. Select maximum effort for every workload without checking whether quality improves.

Explanation

Correct Answer

C:

The appropriate setting depends on measured tradeoffs rather than assuming maximum effort is always best. Measure the workload's quality and latency at supported settings, choosing the lowest-cost configuration that satisfies the required performance.

Incorrect Answers

A:

Additional effort can increase cost or latency without a useful quality gain.

B:

The setting exists to influence how resources are spent and should be evaluated accordingly.

D:

Additional computation may increase cost or latency without a useful gain on simple tasks that already meet targets.

Question 244

Claude is asked for a newly created internal project code that was never supplied or retrieved. What limitation matters?

  1. The project’s name automatically identifies its secret code.
  2. The model lacks authorized evidence for that private new fact.
  3. A confident answer proves prior access.
  4. A larger output allowance reveals any private database.

Explanation

Correct Answer

B:

General training knowledge cannot establish a specific unseen internal value; the application must obtain it from an appropriate source.

Incorrect Answers

A:

Names do not necessarily encode private identifiers.

C:

Fluent text can be unsupported and is not access evidence.

D:

Generation capacity does not grant database access.

Question 245

A prompt says “ignore your context limit.” What effect should an application assume?

  1. No expansion of the model’s actual supported context capacity.
  2. All older messages become permanently stored elsewhere.
  3. The request gains unlimited input capacity.
  4. Only tool definitions stop counting toward context.

Explanation

Correct Answer

A:

A natural-language instruction cannot enlarge a model's technical input capacity; the application must count, select, or summarize context to fit.

Incorrect Answers

B:

No persistence mechanism is created by that sentence.

C:

Capacity is not negotiated by prose.

D:

The instruction does not change token accounting categories.

Question 246

A long identifier is split into several tokens. Does this mean the model receives several separate application identifiers?

  1. Yes; each token becomes a new database key.
  2. No; tokenization is an internal representation of the input sequence.
  3. Yes; the identifier is necessarily invalid.
  4. No; identifiers are always exactly one token.

Explanation

Correct Answer

B:

Application meaning can span token boundaries; tokens are not necessarily whole words or resource identifiers.

Incorrect Answers

A:

Token segmentation does not create external records.

C:

Validity depends on the application contract, not tokenizer segmentation.

D:

Many strings can require multiple tokens, so that universal claim is false.

Question 247

In a simplified autoregressive example, the first desired token has probability 0.5. Given that token, the second desired token has probability 0.2. What is the probability of generating that exact two-token sequence?

  1. 0.35.
  2. 0.20.
  3. 0.10.
  4. 0.70.

Explanation

Correct Answer

C:

The sequence probability is the first token's probability multiplied by the conditional probability of the second: 0.5 times 0.2 equals 0.10.

Incorrect Answers

A:

Averaging the two values measures neither the joint event nor its conditional probability. Exact sequence probability uses multiplication here.

B:

This is only the probability of the second token after the first has occurred. It omits the probability of reaching that first-token condition.

D:

Adding the probabilities does not represent both tokens occurring in sequence. The second probability is conditional on the first token already being selected.

Question 248

Adding many demonstrations improves one metric but exceeds the input budget. What is a reasonable response?

  1. Increase examples indefinitely because more is always better.
  2. Select representative examples with high learning value and measure the tradeoff.
  3. Assume examples do not count as input.
  4. Remove required instructions before reviewing redundant examples.

Explanation

Correct Answer

B:

Examples consume context; choosing informative coverage can preserve guidance without unlimited prompt growth.

Incorrect Answers

A:

Additional examples can add cost, redundancy, or conflicting patterns.

C:

Demonstrations are part of the request context.

D:

Losing the governing task can undermine the examples’ meaning.

Question 249

For a model where thinking consumes output budget, why reserve room beyond the expected final prose?

  1. Reasoning and final answer share the relevant generation allowance.
  2. Final prose is never billed.
  3. Input documents become output tokens automatically.
  4. Thinking automatically enlarges the context window.

Explanation

Correct Answer

A:

A budget sized only for visible prose can leave insufficient room for the complete response under that configuration.

Incorrect Answers

B:

Visible generation remains part of applicable output usage.

C:

Budget categories are not interchangeable simply because thinking is enabled.

D:

Reasoning behavior does not remove the configured capacity boundary.

Question 250

A hypothetical generator succeeds on each independent attempt with probability 0.8. A perfect validator recognizes success, and the system tries at most twice. What is the probability of at least one success?

  1. 1.60.
  2. 0.96.
  3. 1.00.
  4. 0.80.

Explanation

Correct Answer

B:

Both attempts fail with probability 0.2 × 0.2 = 0.04. Subtracting that joint failure from one gives 0.96; the calculation depends on the stated independence and perfect-validator assumptions.

Incorrect Answers

A:

Adding two success probabilities double-counts outcomes where both succeed and exceeds the valid range for a probability.

C:

Two attempts improve the chance of success but do not eliminate the 0.04 probability that both independently fail.

D:

This is the success probability of one attempt. It omits the chance that the second attempt succeeds after the first fails.

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