McAfee-Secured Website

Exam Code: AI-300

Exam Name: Operationalizing Machine Learning and Generative AI Solutions

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Machine Learning Operations Engineer Associate

100% Microsoft AI-300 Guaranteed Success With Testking Training Materials

Get Real AI-300 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

213 Questions & Answers with Testing Engine

"Operationalizing Machine Learning and Generative AI Solutions Exam", also known as AI-300 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-300 Exam Engine. Your AI-300 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-300 Practice Test 55

Question 540

A summarization prompt hard-codes one document type, yet the application accepts several document formats. What design improves reuse?

  1. Parameterize the prompt's document type and task-specific context fields.
  2. Require explicit uncertainty or no-answer behavior when supporting material is unavailable.
  3. Clearly separate untrusted customer content from trusted system-level instructions.
  4. Move authoritative privacy rules into a consistent higher-priority instruction set.

Explanation

Correct Answer

A:

Named prompt variables allow controlled reuse while retaining explicit instructions for each supported input format.

Incorrect Answers

B:

Require explicit uncertainty or no-answer behavior when supporting material is unavailable. This is useful for the distinct issue: A prompt instructs an assistant to invent a reference when source documents are absent. The relevant technical rationale or evidence for that issue is: A clear abstention rule discourages plausible-sounding fabricated citations and unsupported assertions in retrieval-based applications. Here the reported issue is: A summarization prompt hard-codes one document type, yet the application accepts several document formats. The relevant response is: Parameterize the prompt's document type and task-specific context fields. The first result cannot verify this different correction: Named prompt variables allow controlled reuse while retaining explicit instructions for each supported input format.

C:

Clearly separate untrusted customer content from trusted system-level instructions. It addresses another case: A prompt includes a customer message containing the text ignore previous instructions, and the model treats it as policy. For that case, the distinguishing justification or acceptance check is: Instruction hierarchy and data delimiters reduce the chance that a retrieved or user-supplied string overrides the application's policy. The present case instead states: A summarization prompt hard-codes one document type, yet the application accepts several document formats. It calls for Parameterize the prompt's document type and task-specific context fields. Showing success for the other case would not establish: Named prompt variables allow controlled reuse while retaining explicit instructions for each supported input format.

D:

The alternative, Move authoritative privacy rules into a consistent higher-priority instruction set, belongs to a different technical condition: A prompt repeats conflicting instructions about privacy in system and user messages. There the source-specific mechanism or proof is: Conflicting instruction placement makes responses unpredictable; an explicit hierarchy protects stable application requirements. The condition asked about here is: A summarization prompt hard-codes one document type, yet the application accepts several document formats. A suitable correction is Parameterize the prompt's document type and task-specific context fields; the targeted justification or result is: Named prompt variables allow controlled reuse while retaining explicit instructions for each supported input format.

Question 541

A model responds to an ambiguous request by fabricating a missing customer identifier rather than asking a question. What should the prompt require?

  1. Define the output schema and constrain responses to the required fields.
  2. Move authoritative privacy rules into a consistent higher-priority instruction set.
  3. Clearly separate untrusted customer content from trusted system-level instructions.
  4. Ask for essential missing inputs instead of guessing protected identifiers.

Explanation

Correct Answer

D:

Clarification rules reduce unauthorized assumptions and prevent unsafe model actions when critical parameters are absent.

Incorrect Answers

A:

Define the output schema and constrain responses to the required fields. A separate source case supports that approach when: A customer-service prompt produces long essays when the application expects a three-field JSON object. The supporting rationale or observable result is: Explicit structured output requirements are more reliable than leaving format assumptions to free-form generation. This question describes a different issue: A model responds to an ambiguous request by fabricating a missing customer identifier rather than asking a question. Address it through Ask for essential missing inputs instead of guessing protected identifiers, with the required justification or evidence: Clarification rules reduce unauthorized assumptions and prevent unsafe model actions when critical parameters are absent.

B:

Move authoritative privacy rules into a consistent higher-priority instruction set. This is useful for the distinct issue: A prompt repeats conflicting instructions about privacy in system and user messages. The relevant technical rationale or evidence for that issue is: Conflicting instruction placement makes responses unpredictable; an explicit hierarchy protects stable application requirements. Here the reported issue is: A model responds to an ambiguous request by fabricating a missing customer identifier rather than asking a question. The relevant response is: Ask for essential missing inputs instead of guessing protected identifiers. The first result cannot verify this different correction: Clarification rules reduce unauthorized assumptions and prevent unsafe model actions when critical parameters are absent.

C:

Clearly separate untrusted customer content from trusted system-level instructions. It addresses another case: A prompt includes a customer message containing the text ignore previous instructions, and the model treats it as policy. For that case, the distinguishing justification or acceptance check is: Instruction hierarchy and data delimiters reduce the chance that a retrieved or user-supplied string overrides the application's policy. The present case instead states: A model responds to an ambiguous request by fabricating a missing customer identifier rather than asking a question. It calls for Ask for essential missing inputs instead of guessing protected identifiers. Showing success for the other case would not establish: Clarification rules reduce unauthorized assumptions and prevent unsafe model actions when critical parameters are absent.

Question 542

An evaluation shows inconsistent outputs because the prompt mixes system policy, retrieved documents, and sample replies into one unstructured block. What is the redesign?

  1. Move authoritative privacy rules into a consistent higher-priority instruction set.
  2. Include representative few-shot input-output examples for the extraction task.
  3. Organize prompt sections into clear policy, context, examples, and current task boundaries.
  4. Delimit quoted source material and label it as data rather than instructions.

Explanation

Correct Answer

C:

Role and section structure makes precedence and evidence sources clearer, improving repeatable task execution.

Incorrect Answers

A:

Move authoritative privacy rules into a consistent higher-priority instruction set. That operation would be relevant if: A prompt repeats conflicting instructions about privacy in system and user messages. Its corresponding reason or verification is: Conflicting instruction placement makes responses unpredictable; an explicit hierarchy protects stable application requirements. However the problem in this question is: An evaluation shows inconsistent outputs because the prompt mixes system policy, retrieved documents, and sample replies into one unstructured block. The requirement instead calls for Organize prompt sections into clear policy, context, examples, and current task boundaries; the question-specific check is: Role and section structure makes precedence and evidence sources clearer, improving repeatable task execution.

B:

Include representative few-shot input-output examples for the extraction task. A separate source case supports that approach when: A prompt needs to extract the same fields from many differently worded invoices. The supporting rationale or observable result is: Examples demonstrate the intended mapping and format beyond vague natural-language directions alone. This question describes a different issue: An evaluation shows inconsistent outputs because the prompt mixes system policy, retrieved documents, and sample replies into one unstructured block. Address it through Organize prompt sections into clear policy, context, examples, and current task boundaries, with the required justification or evidence: Role and section structure makes precedence and evidence sources clearer, improving repeatable task execution.

D:

Delimit quoted source material and label it as data rather than instructions. This is useful for the distinct issue: A model has to distinguish quoted legal text from the question asked about it, but both appear as unmarked prose. The relevant technical rationale or evidence for that issue is: Clear delimiters reduce confusion between the content to analyze and the instructions controlling analysis. Here the reported issue is: An evaluation shows inconsistent outputs because the prompt mixes system policy, retrieved documents, and sample replies into one unstructured block. The relevant response is: Organize prompt sections into clear policy, context, examples, and current task boundaries. The first result cannot verify this different correction: Role and section structure makes precedence and evidence sources clearer, improving repeatable task execution.

Question 543

A prompt engineer changes both retrieval context and instructions in one test, then attributes a quality improvement to wording alone. How should prompt variants be compared?

  1. Inspect disagreement and failure cases with human or documented AI-assisted adjudication.
  2. Use representative test cases covering normal, edge, and high-risk requests.
  3. Change one controlled factor at a time or use a factorial evaluation design.
  4. Compare quality improvements against tokens, latency, and unit economics.

Explanation

Correct Answer

C:

Controlled experiments distinguish prompt wording effects from changes to retrieved evidence and model settings.

Incorrect Answers

A:

The alternative, Inspect disagreement and failure cases with human or documented AI-assisted adjudication, belongs to a different technical condition: A team compares prompts only through automated judge scores, with no investigation of obvious contradictory cases. There the source-specific mechanism or proof is: Evaluator averages cannot replace diagnosis of specific ambiguous or potentially unsafe outputs. The condition asked about here is: A prompt engineer changes both retrieval context and instructions in one test, then attributes a quality improvement to wording alone. A suitable correction is Change one controlled factor at a time or use a factorial evaluation design; the targeted justification or result is: Controlled experiments distinguish prompt wording effects from changes to retrieved evidence and model settings.

B:

Use representative test cases covering normal, edge, and high-risk requests. That operation would be relevant if: A prompt variant performs well on a dozen hand-selected easy cases but fails on real support traffic. Its corresponding reason or verification is: A biased test set can make a prompt look successful without demonstrating production suitability. However the problem in this question is: A prompt engineer changes both retrieval context and instructions in one test, then attributes a quality improvement to wording alone. The requirement instead calls for Change one controlled factor at a time or use a factorial evaluation design; the question-specific check is: Controlled experiments distinguish prompt wording effects from changes to retrieved evidence and model settings.

D:

Compare quality improvements against tokens, latency, and unit economics. A separate source case supports that approach when: A prompt rollout improves customer ratings but doubles token costs. The supporting rationale or observable result is: A variant should meet both quality requirements and operational resource budgets before production promotion. This question describes a different issue: A prompt engineer changes both retrieval context and instructions in one test, then attributes a quality improvement to wording alone. Address it through Change one controlled factor at a time or use a factorial evaluation design, with the required justification or evidence: Controlled experiments distinguish prompt wording effects from changes to retrieved evidence and model settings.

Question 544

Two prompt candidates are tested on unrelated user questions, and the team compares average accuracy. What setup improves validity?

  1. Separate retrieval-version and prompt-version changes in a controlled comparison.
  2. Evaluate both variants against the same representative labeled evaluation cases.
  3. Validate the selected prompt on a separate untouched test set.
  4. Include latency percentiles and failure rates alongside user quality metrics.

Explanation

Correct Answer

B:

Paired inputs make outcome differences attributable to the prompt rather than changing test-set difficulty.

Incorrect Answers

A:

Separate retrieval-version and prompt-version changes in a controlled comparison. It addresses another case: An evaluator returns an improved groundedness score for a new prompt but the retrieval service was also upgraded. For that case, the distinguishing justification or acceptance check is: Without isolating both factors, the experiment cannot identify which modification caused the observed improvement. The present case instead states: Two prompt candidates are tested on unrelated user questions, and the team compares average accuracy. It calls for Evaluate both variants against the same representative labeled evaluation cases. Showing success for the other case would not establish: Paired inputs make outcome differences attributable to the prompt rather than changing test-set difficulty.

C:

The alternative, Validate the selected prompt on a separate untouched test set, belongs to a different technical condition: An evaluation selects the best prompt after repeatedly tuning against the same 20 examples. There the source-specific mechanism or proof is: Repeated tuning can memorize evaluation examples; an independent holdout provides more realistic generalization evidence. The condition asked about here is: Two prompt candidates are tested on unrelated user questions, and the team compares average accuracy. A suitable correction is Evaluate both variants against the same representative labeled evaluation cases; the targeted justification or result is: Paired inputs make outcome differences attributable to the prompt rather than changing test-set difficulty.

D:

Include latency percentiles and failure rates alongside user quality metrics. That operation would be relevant if: Customer satisfaction improves under prompt B, but response times worsen beyond the app's SLA. Its corresponding reason or verification is: Production acceptance depends on user experience constraints beyond subjective response quality. However the problem in this question is: Two prompt candidates are tested on unrelated user questions, and the team compares average accuracy. The requirement instead calls for Evaluate both variants against the same representative labeled evaluation cases; the question-specific check is: Paired inputs make outcome differences attributable to the prompt rather than changing test-set difficulty.

Question 545

A new instruction improves average fluency but hurts groundedness on sensitive policies. What should the acceptance criteria require?

  1. Compare both fluency and groundedness against release thresholds.
  2. Use a broad regression dataset weighted to real workload distribution and critical edge cases.
  3. Use representative test cases covering normal, edge, and high-risk requests.
  4. Include latency percentiles and failure rates alongside user quality metrics.

Explanation

Correct Answer

A:

Prompt quality is multidimensional; one attractive metric cannot override an important factual reliability regression.

Incorrect Answers

B:

Use a broad regression dataset weighted to real workload distribution and critical edge cases. This is useful for the distinct issue: A prompt candidate fixes one high-profile complaint but degrades several routine intents. The relevant technical rationale or evidence for that issue is: An isolated success case is insufficient evidence that overall prompt behavior remains acceptable. Here the reported issue is: A new instruction improves average fluency but hurts groundedness on sensitive policies. The relevant response is: Compare both fluency and groundedness against release thresholds. The first result cannot verify this different correction: Prompt quality is multidimensional; one attractive metric cannot override an important factual reliability regression.

C:

Use representative test cases covering normal, edge, and high-risk requests. It addresses another case: A prompt variant performs well on a dozen hand-selected easy cases but fails on real support traffic. For that case, the distinguishing justification or acceptance check is: A biased test set can make a prompt look successful without demonstrating production suitability. The present case instead states: A new instruction improves average fluency but hurts groundedness on sensitive policies. It calls for Compare both fluency and groundedness against release thresholds. Showing success for the other case would not establish: Prompt quality is multidimensional; one attractive metric cannot override an important factual reliability regression.

D:

The alternative, Include latency percentiles and failure rates alongside user quality metrics, belongs to a different technical condition: Customer satisfaction improves under prompt B, but response times worsen beyond the app's SLA. There the source-specific mechanism or proof is: Production acceptance depends on user experience constraints beyond subjective response quality. The condition asked about here is: A new instruction improves average fluency but hurts groundedness on sensitive policies. A suitable correction is Compare both fluency and groundedness against release thresholds; the targeted justification or result is: Prompt quality is multidimensional; one attractive metric cannot override an important factual reliability regression.

Question 546

A prompt A/B test assigns only enterprise users to variant A and only guests to B. What design correction is needed?

  1. Inspect disagreement and failure cases with human or documented AI-assisted adjudication.
  2. Stage the prompt change through a controlled canary with rollback capability.
  3. Include latency percentiles and failure rates alongside user quality metrics.
  4. Randomize variant assignment within appropriate comparable user cohorts.

Explanation

Correct Answer

D:

Nonrandom audiences can confound prompt performance differences with distinct user populations and request patterns.

Incorrect Answers

A:

Inspect disagreement and failure cases with human or documented AI-assisted adjudication. A separate source case supports that approach when: A team compares prompts only through automated judge scores, with no investigation of obvious contradictory cases. The supporting rationale or observable result is: Evaluator averages cannot replace diagnosis of specific ambiguous or potentially unsafe outputs. This question describes a different issue: A prompt A/B test assigns only enterprise users to variant A and only guests to B. Address it through Randomize variant assignment within appropriate comparable user cohorts, with the required justification or evidence: Nonrandom audiences can confound prompt performance differences with distinct user populations and request patterns.

B:

Stage the prompt change through a controlled canary with rollback capability. This is useful for the distinct issue: A prompt candidate is deployed to all users after a lab test, and unexpected production wording issues emerge. The relevant technical rationale or evidence for that issue is: A limited rollout detects population-specific regressions while retaining the prior prompt for most users. Here the reported issue is: A prompt A/B test assigns only enterprise users to variant A and only guests to B. The relevant response is: Randomize variant assignment within appropriate comparable user cohorts. The first result cannot verify this different correction: Nonrandom audiences can confound prompt performance differences with distinct user populations and request patterns.

C:

Include latency percentiles and failure rates alongside user quality metrics. It addresses another case: Customer satisfaction improves under prompt B, but response times worsen beyond the app's SLA. For that case, the distinguishing justification or acceptance check is: Production acceptance depends on user experience constraints beyond subjective response quality. The present case instead states: A prompt A/B test assigns only enterprise users to variant A and only guests to B. It calls for Randomize variant assignment within appropriate comparable user cohorts. Showing success for the other case would not establish: Nonrandom audiences can confound prompt performance differences with distinct user populations and request patterns.

Question 547

A prompt variant performs well on a dozen hand-selected easy cases but fails on real support traffic. What should evaluation sampling improve?

  1. Stage the prompt change through a controlled canary with rollback capability.
  2. Include prompt-injection and untrusted-context cases in the variant test set.
  3. Use representative test cases covering normal, edge, and high-risk requests.
  4. Validate the selected prompt on a separate untouched test set.

Explanation

Correct Answer

C:

A biased test set can make a prompt look successful without demonstrating production suitability.

Incorrect Answers

A:

Stage the prompt change through a controlled canary with rollback capability. That operation would be relevant if: A prompt candidate is deployed to all users after a lab test, and unexpected production wording issues emerge. Its corresponding reason or verification is: A limited rollout detects population-specific regressions while retaining the prior prompt for most users. However the problem in this question is: A prompt variant performs well on a dozen hand-selected easy cases but fails on real support traffic. The requirement instead calls for Use representative test cases covering normal, edge, and high-risk requests; the question-specific check is: A biased test set can make a prompt look successful without demonstrating production suitability.

B:

Include prompt-injection and untrusted-context cases in the variant test set. A separate source case supports that approach when: A new prompt's evaluation dataset omits adversarial instructions embedded in retrieved documents. The supporting rationale or observable result is: Security-relevant prompt changes need adversarial coverage, not only ordinary benign user questions. This question describes a different issue: A prompt variant performs well on a dozen hand-selected easy cases but fails on real support traffic. Address it through Use representative test cases covering normal, edge, and high-risk requests, with the required justification or evidence: A biased test set can make a prompt look successful without demonstrating production suitability.

D:

Validate the selected prompt on a separate untouched test set. This is useful for the distinct issue: An evaluation selects the best prompt after repeatedly tuning against the same 20 examples. The relevant technical rationale or evidence for that issue is: Repeated tuning can memorize evaluation examples; an independent holdout provides more realistic generalization evidence. Here the reported issue is: A prompt variant performs well on a dozen hand-selected easy cases but fails on real support traffic. The relevant response is: Use representative test cases covering normal, edge, and high-risk requests. The first result cannot verify this different correction: A biased test set can make a prompt look successful without demonstrating production suitability.

Question 548

A prompt evaluator changes the model version halfway through a comparison and still attributes every metric change to the prompt. What must be held constant?

  1. Validate the selected prompt on a separate untouched test set.
  2. Pin the underlying model deployment and test configuration for the variant experiment.
  3. Use representative test cases covering normal, edge, and high-risk requests.
  4. Compare quality improvements against tokens, latency, and unit economics.

Explanation

Correct Answer

B:

A model change is a confounding variable that invalidates claims about prompt-only performance improvements.

Incorrect Answers

A:

The alternative, Validate the selected prompt on a separate untouched test set, belongs to a different technical condition: An evaluation selects the best prompt after repeatedly tuning against the same 20 examples. There the source-specific mechanism or proof is: Repeated tuning can memorize evaluation examples; an independent holdout provides more realistic generalization evidence. The condition asked about here is: A prompt evaluator changes the model version halfway through a comparison and still attributes every metric change to the prompt. A suitable correction is Pin the underlying model deployment and test configuration for the variant experiment; the targeted justification or result is: A model change is a confounding variable that invalidates claims about prompt-only performance improvements.

C:

Use representative test cases covering normal, edge, and high-risk requests. That operation would be relevant if: A prompt variant performs well on a dozen hand-selected easy cases but fails on real support traffic. Its corresponding reason or verification is: A biased test set can make a prompt look successful without demonstrating production suitability. However the problem in this question is: A prompt evaluator changes the model version halfway through a comparison and still attributes every metric change to the prompt. The requirement instead calls for Pin the underlying model deployment and test configuration for the variant experiment; the question-specific check is: A model change is a confounding variable that invalidates claims about prompt-only performance improvements.

D:

Compare quality improvements against tokens, latency, and unit economics. A separate source case supports that approach when: A prompt rollout improves customer ratings but doubles token costs. The supporting rationale or observable result is: A variant should meet both quality requirements and operational resource budgets before production promotion. This question describes a different issue: A prompt evaluator changes the model version halfway through a comparison and still attributes every metric change to the prompt. Address it through Pin the underlying model deployment and test configuration for the variant experiment, with the required justification or evidence: A model change is a confounding variable that invalidates claims about prompt-only performance improvements.

Question 549

A prompt rollout improves customer ratings but doubles token costs. What should the evaluation include?

  1. Compare quality improvements against tokens, latency, and unit economics.
  2. Pin the underlying model deployment and test configuration for the variant experiment.
  3. Compare both fluency and groundedness against release thresholds.
  4. Include prompt-injection and untrusted-context cases in the variant test set.

Explanation

Correct Answer

A:

A variant should meet both quality requirements and operational resource budgets before production promotion.

Incorrect Answers

B:

Pin the underlying model deployment and test configuration for the variant experiment. It addresses another case: A prompt evaluator changes the model version halfway through a comparison and still attributes every metric change to the prompt. For that case, the distinguishing justification or acceptance check is: A model change is a confounding variable that invalidates claims about prompt-only performance improvements. The present case instead states: A prompt rollout improves customer ratings but doubles token costs. It calls for Compare quality improvements against tokens, latency, and unit economics. Showing success for the other case would not establish: A variant should meet both quality requirements and operational resource budgets before production promotion.

C:

The alternative, Compare both fluency and groundedness against release thresholds, belongs to a different technical condition: A new instruction improves average fluency but hurts groundedness on sensitive policies. There the source-specific mechanism or proof is: Prompt quality is multidimensional; one attractive metric cannot override an important factual reliability regression. The condition asked about here is: A prompt rollout improves customer ratings but doubles token costs. A suitable correction is Compare quality improvements against tokens, latency, and unit economics; the targeted justification or result is: A variant should meet both quality requirements and operational resource budgets before production promotion.

D:

Include prompt-injection and untrusted-context cases in the variant test set. That operation would be relevant if: A new prompt's evaluation dataset omits adversarial instructions embedded in retrieved documents. Its corresponding reason or verification is: Security-relevant prompt changes need adversarial coverage, not only ordinary benign user questions. However the problem in this question is: A prompt rollout improves customer ratings but doubles token costs. The requirement instead calls for Compare quality improvements against tokens, latency, and unit economics; the question-specific check is: A variant should meet both quality requirements and operational resource budgets before production promotion.

Question 550

An evaluation selects the best prompt after repeatedly tuning against the same 20 examples. What guards against overfitting?

  1. Include prompt-injection and untrusted-context cases in the variant test set.
  2. Use a broad regression dataset weighted to real workload distribution and critical edge cases.
  3. Evaluate both variants against the same representative labeled evaluation cases.
  4. Validate the selected prompt on a separate untouched test set.

Explanation

Correct Answer

D:

Repeated tuning can memorize evaluation examples; an independent holdout provides more realistic generalization evidence.

Incorrect Answers

A:

Include prompt-injection and untrusted-context cases in the variant test set. It addresses another case: A new prompt's evaluation dataset omits adversarial instructions embedded in retrieved documents. For that case, the distinguishing justification or acceptance check is: Security-relevant prompt changes need adversarial coverage, not only ordinary benign user questions. The present case instead states: An evaluation selects the best prompt after repeatedly tuning against the same 20 examples. It calls for Validate the selected prompt on a separate untouched test set. Showing success for the other case would not establish: Repeated tuning can memorize evaluation examples; an independent holdout provides more realistic generalization evidence.

B:

The alternative, Use a broad regression dataset weighted to real workload distribution and critical edge cases, belongs to a different technical condition: A prompt candidate fixes one high-profile complaint but degrades several routine intents. There the source-specific mechanism or proof is: An isolated success case is insufficient evidence that overall prompt behavior remains acceptable. The condition asked about here is: An evaluation selects the best prompt after repeatedly tuning against the same 20 examples. A suitable correction is Validate the selected prompt on a separate untouched test set; the targeted justification or result is: Repeated tuning can memorize evaluation examples; an independent holdout provides more realistic generalization evidence.

C:

Evaluate both variants against the same representative labeled evaluation cases. That operation would be relevant if: Two prompt candidates are tested on unrelated user questions, and the team compares average accuracy. Its corresponding reason or verification is: Paired inputs make outcome differences attributable to the prompt rather than changing test-set difficulty. However the problem in this question is: An evaluation selects the best prompt after repeatedly tuning against the same 20 examples. The requirement instead calls for Validate the selected prompt on a separate untouched test set; the question-specific check is: Repeated tuning can memorize evaluation examples; an independent holdout provides more realistic generalization evidence.

Operationalizing Machine Learning and Generative AI Solutions (AI-300) Practice Tests