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 58

Question 570

A project contains many large prompt-evaluation datasets inside the same Git repository, inflating clone times. What should be separated?

  1. Create an approved release tag for the exact prompt commit.
  2. Pin workflow actions to reviewed commit SHAs and minimize token scopes.
  3. Record and deploy the exact prompt Git commit SHA or immutable release tag.
  4. Keep large governed evaluation datasets in versioned storage and reference them from Git metadata.

Explanation

Correct Answer

D:

Git tracks prompt source effectively; external governed data versioning handles bulky test corpora more safely.

Incorrect Answers

A:

Create an approved release tag for the exact prompt commit. That operation would be relevant if: The organization wants a traceable release named prompts-v2.4 across several services. Its corresponding reason or verification is: A protected tag can identify the immutable prompt revision all consumers should deploy and audit. However the problem in this question is: A project contains many large prompt-evaluation datasets inside the same Git repository, inflating clone times. The requirement instead calls for Keep large governed evaluation datasets in versioned storage and reference them from Git metadata; the question-specific check is: Git tracks prompt source effectively; external governed data versioning handles bulky test corpora more safely.

B:

Pin workflow actions to reviewed commit SHAs and minimize token scopes. A separate source case supports that approach when: An automatic prompt deployment downloads actions from unpinned third-party GitHub repositories with broad token permissions. The supporting rationale or observable result is: Reproducible dependencies and least privilege reduce the chance that automation changes unexpectedly during a sensitive release. This question describes a different issue: A project contains many large prompt-evaluation datasets inside the same Git repository, inflating clone times. Address it through Keep large governed evaluation datasets in versioned storage and reference them from Git metadata, with the required justification or evidence: Git tracks prompt source effectively; external governed data versioning handles bulky test corpora more safely.

C:

Record and deploy the exact prompt Git commit SHA or immutable release tag. This is useful for the distinct issue: A deployed prompt is referenced as main branch latest, and reviewers cannot reproduce yesterday's responses. The relevant technical rationale or evidence for that issue is: Immutable revisions identify the instructions used by each release independently of subsequent branch changes. Here the reported issue is: A project contains many large prompt-evaluation datasets inside the same Git repository, inflating clone times. The relevant response is: Keep large governed evaluation datasets in versioned storage and reference them from Git metadata. The first result cannot verify this different correction: Git tracks prompt source effectively; external governed data versioning handles bulky test corpora more safely.

Question 571

The organization wants a traceable release named prompts-v2.4 across several services. What Git feature can provide a stable marker?

  1. Store production prompt templates as version-controlled files in Git.
  2. Revert the change through a new reviewed Git commit and redeploy.
  3. Create an approved release tag for the exact prompt commit.
  4. Attach the prompt evaluation summary and failing cases to the pull request or release record.

Explanation

Correct Answer

C:

A protected tag can identify the immutable prompt revision all consumers should deploy and audit.

Incorrect Answers

A:

The alternative, Store production prompt templates as version-controlled files in Git, belongs to a different technical condition: A production agent loads its prompt from a shared editable text file and behavior changes without review. There the source-specific mechanism or proof is: Source control creates traceable changes and enables exact prompt restoration when an unexpected regression occurs. The condition asked about here is: The organization wants a traceable release named prompts-v2.4 across several services. A suitable correction is Create an approved release tag for the exact prompt commit; the targeted justification or result is: A protected tag can identify the immutable prompt revision all consumers should deploy and audit.

B:

Revert the change through a new reviewed Git commit and redeploy. That operation would be relevant if: A prompt change accidentally causes unsafe tool calls, and operations need to restore known behavior without deleting audit history. Its corresponding reason or verification is: An auditable revert preserves the change history while restoring the earlier prompt behavior. However the problem in this question is: The organization wants a traceable release named prompts-v2.4 across several services. The requirement instead calls for Create an approved release tag for the exact prompt commit; the question-specific check is: A protected tag can identify the immutable prompt revision all consumers should deploy and audit.

D:

Attach the prompt evaluation summary and failing cases to the pull request or release record. A separate source case supports that approach when: A developer edits a prompt and runs ad hoc tests, but none of those results appear in source review. The supporting rationale or observable result is: Reviewers need visible behavioral evidence to judge prompt changes, not only textual diffs in Git. This question describes a different issue: The organization wants a traceable release named prompts-v2.4 across several services. Address it through Create an approved release tag for the exact prompt commit, with the required justification or evidence: A protected tag can identify the immutable prompt revision all consumers should deploy and audit.

Question 572

A developer edits a prompt and runs ad hoc tests, but none of those results appear in source review. What should CI retain?

  1. Compare the deployed prompt artifact hash against the approved Git release artifact.
  2. Attach the prompt evaluation summary and failing cases to the pull request or release record.
  3. Include prompt commit, model deployment version, tools, and evaluation dataset versions.
  4. Remove the credential from prompt source and use a protected tool secret store.

Explanation

Correct Answer

B:

Reviewers need visible behavioral evidence to judge prompt changes, not only textual diffs in Git.

Incorrect Answers

A:

Compare the deployed prompt artifact hash against the approved Git release artifact. It addresses another case: Security reviewers want proof that the current production prompt still matches its approved repository revision. For that case, the distinguishing justification or acceptance check is: Hash verification detects manual or accidental changes to model-visible instructions after review and promotion. The present case instead states: A developer edits a prompt and runs ad hoc tests, but none of those results appear in source review. It calls for Attach the prompt evaluation summary and failing cases to the pull request or release record. Showing success for the other case would not establish: Reviewers need visible behavioral evidence to judge prompt changes, not only textual diffs in Git.

C:

The alternative, Include prompt commit, model deployment version, tools, and evaluation dataset versions, belongs to a different technical condition: A rollback requires restoring a prompt plus the matching model version, but Git tracks only prompt text. There the source-specific mechanism or proof is: Model and prompt combinations jointly determine behavior; recovering only one may not reproduce the previous release. The condition asked about here is: A developer edits a prompt and runs ad hoc tests, but none of those results appear in source review. A suitable correction is Attach the prompt evaluation summary and failing cases to the pull request or release record; the targeted justification or result is: Reviewers need visible behavioral evidence to judge prompt changes, not only textual diffs in Git.

D:

Remove the credential from prompt source and use a protected tool secret store. That operation would be relevant if: A production prompt contains an embedded third-party API key that is visible to every repository contributor. Its corresponding reason or verification is: Prompts can be logged, inspected, or echoed; API secrets belong in controlled infrastructure, not text instructions. However the problem in this question is: A developer edits a prompt and runs ad hoc tests, but none of those results appear in source review. The requirement instead calls for Attach the prompt evaluation summary and failing cases to the pull request or release record; the question-specific check is: Reviewers need visible behavioral evidence to judge prompt changes, not only textual diffs in Git.

Question 573

An automatic prompt deployment downloads actions from unpinned third-party GitHub repositories with broad token permissions. What supply-chain control helps?

  1. Pin workflow actions to reviewed commit SHAs and minimize token scopes.
  2. Block prompt deployment unless required evaluation and regression checks pass.
  3. Create an approved release tag for the exact prompt commit.
  4. Include prompt commit, model deployment version, tools, and evaluation dataset versions.

Explanation

Correct Answer

A:

Reproducible dependencies and least privilege reduce the chance that automation changes unexpectedly during a sensitive release.

Incorrect Answers

B:

Block prompt deployment unless required evaluation and regression checks pass. This is useful for the distinct issue: A prompt evaluation fails after an update, but a CI pipeline still publishes the changed prompt. The relevant technical rationale or evidence for that issue is: CI status gates prevent an unsafe or broken instruction change from reaching the production application. Here the reported issue is: An automatic prompt deployment downloads actions from unpinned third-party GitHub repositories with broad token permissions. The relevant response is: Pin workflow actions to reviewed commit SHAs and minimize token scopes. The first result cannot verify this different correction: Reproducible dependencies and least privilege reduce the chance that automation changes unexpectedly during a sensitive release.

C:

Create an approved release tag for the exact prompt commit. It addresses another case: The organization wants a traceable release named prompts-v2.4 across several services. For that case, the distinguishing justification or acceptance check is: A protected tag can identify the immutable prompt revision all consumers should deploy and audit. The present case instead states: An automatic prompt deployment downloads actions from unpinned third-party GitHub repositories with broad token permissions. It calls for Pin workflow actions to reviewed commit SHAs and minimize token scopes. Showing success for the other case would not establish: Reproducible dependencies and least privilege reduce the chance that automation changes unexpectedly during a sensitive release.

D:

The alternative, Include prompt commit, model deployment version, tools, and evaluation dataset versions, belongs to a different technical condition: A rollback requires restoring a prompt plus the matching model version, but Git tracks only prompt text. There the source-specific mechanism or proof is: Model and prompt combinations jointly determine behavior; recovering only one may not reproduce the previous release. The condition asked about here is: An automatic prompt deployment downloads actions from unpinned third-party GitHub repositories with broad token permissions. A suitable correction is Pin workflow actions to reviewed commit SHAs and minimize token scopes; the targeted justification or result is: Reproducible dependencies and least privilege reduce the chance that automation changes unexpectedly during a sensitive release.

Question 574

A prompt template contains different production and staging API addresses hard-coded into separate file copies. What should be changed?

  1. Promote the same immutable reviewed prompt artifact across environments.
  2. Store production prompt templates as version-controlled files in Git.
  3. Compare the deployed prompt artifact hash against the approved Git release artifact.
  4. Parameterize environment-specific references while versioning one common prompt template.

Explanation

Correct Answer

D:

Parameterization preserves consistent instructions without embedding configuration drift in copied prompt source files.

Incorrect Answers

A:

Promote the same immutable reviewed prompt artifact across environments. A separate source case supports that approach when: A model team manually copies prompt files between staging and production, and one file differs by a missing line. The supporting rationale or observable result is: Artifact promotion preserves exact instructions without retyping or manually recreating a purportedly identical production file. This question describes a different issue: A prompt template contains different production and staging API addresses hard-coded into separate file copies. Address it through Parameterize environment-specific references while versioning one common prompt template, with the required justification or evidence: Parameterization preserves consistent instructions without embedding configuration drift in copied prompt source files.

B:

Store production prompt templates as version-controlled files in Git. This is useful for the distinct issue: A production agent loads its prompt from a shared editable text file and behavior changes without review. The relevant technical rationale or evidence for that issue is: Source control creates traceable changes and enables exact prompt restoration when an unexpected regression occurs. Here the reported issue is: A prompt template contains different production and staging API addresses hard-coded into separate file copies. The relevant response is: Parameterize environment-specific references while versioning one common prompt template. The first result cannot verify this different correction: Parameterization preserves consistent instructions without embedding configuration drift in copied prompt source files.

C:

Compare the deployed prompt artifact hash against the approved Git release artifact. It addresses another case: Security reviewers want proof that the current production prompt still matches its approved repository revision. For that case, the distinguishing justification or acceptance check is: Hash verification detects manual or accidental changes to model-visible instructions after review and promotion. The present case instead states: A prompt template contains different production and staging API addresses hard-coded into separate file copies. It calls for Parameterize environment-specific references while versioning one common prompt template. Showing success for the other case would not establish: Parameterization preserves consistent instructions without embedding configuration drift in copied prompt source files.

Question 575

A rollback requires restoring a prompt plus the matching model version, but Git tracks only prompt text. What should a release manifest record?

  1. Block prompt deployment unless required evaluation and regression checks pass.
  2. Pin workflow actions to reviewed commit SHAs and minimize token scopes.
  3. Include prompt commit, model deployment version, tools, and evaluation dataset versions.
  4. Record and deploy the exact prompt Git commit SHA or immutable release tag.

Explanation

Correct Answer

C:

Model and prompt combinations jointly determine behavior; recovering only one may not reproduce the previous release.

Incorrect Answers

A:

Block prompt deployment unless required evaluation and regression checks pass. That operation would be relevant if: A prompt evaluation fails after an update, but a CI pipeline still publishes the changed prompt. Its corresponding reason or verification is: CI status gates prevent an unsafe or broken instruction change from reaching the production application. However the problem in this question is: A rollback requires restoring a prompt plus the matching model version, but Git tracks only prompt text. The requirement instead calls for Include prompt commit, model deployment version, tools, and evaluation dataset versions; the question-specific check is: Model and prompt combinations jointly determine behavior; recovering only one may not reproduce the previous release.

B:

Pin workflow actions to reviewed commit SHAs and minimize token scopes. A separate source case supports that approach when: An automatic prompt deployment downloads actions from unpinned third-party GitHub repositories with broad token permissions. The supporting rationale or observable result is: Reproducible dependencies and least privilege reduce the chance that automation changes unexpectedly during a sensitive release. This question describes a different issue: A rollback requires restoring a prompt plus the matching model version, but Git tracks only prompt text. Address it through Include prompt commit, model deployment version, tools, and evaluation dataset versions, with the required justification or evidence: Model and prompt combinations jointly determine behavior; recovering only one may not reproduce the previous release.

D:

Record and deploy the exact prompt Git commit SHA or immutable release tag. This is useful for the distinct issue: A deployed prompt is referenced as main branch latest, and reviewers cannot reproduce yesterday's responses. The relevant technical rationale or evidence for that issue is: Immutable revisions identify the instructions used by each release independently of subsequent branch changes. Here the reported issue is: A rollback requires restoring a prompt plus the matching model version, but Git tracks only prompt text. The relevant response is: Include prompt commit, model deployment version, tools, and evaluation dataset versions. The first result cannot verify this different correction: Model and prompt combinations jointly determine behavior; recovering only one may not reproduce the previous release.

Question 576

Security reviewers want proof that the current production prompt still matches its approved repository revision. What should deployment automation verify?

  1. Require reviewed pull requests and passing prompt tests on the protected branch.
  2. Compare the deployed prompt artifact hash against the approved Git release artifact.
  3. Revert the change through a new reviewed Git commit and redeploy.
  4. Remove the credential from prompt source and use a protected tool secret store.

Explanation

Correct Answer

B:

Hash verification detects manual or accidental changes to model-visible instructions after review and promotion.

Incorrect Answers

A:

The alternative, Require reviewed pull requests and passing prompt tests on the protected branch, belongs to a different technical condition: A prompt author pushes changed safety instructions directly to main and the agent deploys automatically. There the source-specific mechanism or proof is: Reviewed merges prevent unchecked behavior changes from entering a production model application. The condition asked about here is: Security reviewers want proof that the current production prompt still matches its approved repository revision. A suitable correction is Compare the deployed prompt artifact hash against the approved Git release artifact; the targeted justification or result is: Hash verification detects manual or accidental changes to model-visible instructions after review and promotion.

C:

Revert the change through a new reviewed Git commit and redeploy. That operation would be relevant if: A prompt change accidentally causes unsafe tool calls, and operations need to restore known behavior without deleting audit history. Its corresponding reason or verification is: An auditable revert preserves the change history while restoring the earlier prompt behavior. However the problem in this question is: Security reviewers want proof that the current production prompt still matches its approved repository revision. The requirement instead calls for Compare the deployed prompt artifact hash against the approved Git release artifact; the question-specific check is: Hash verification detects manual or accidental changes to model-visible instructions after review and promotion.

D:

Remove the credential from prompt source and use a protected tool secret store. A separate source case supports that approach when: A production prompt contains an embedded third-party API key that is visible to every repository contributor. The supporting rationale or observable result is: Prompts can be logged, inspected, or echoed; API secrets belong in controlled infrastructure, not text instructions. This question describes a different issue: Security reviewers want proof that the current production prompt still matches its approved repository revision. Address it through Compare the deployed prompt artifact hash against the approved Git release artifact, with the required justification or evidence: Hash verification detects manual or accidental changes to model-visible instructions after review and promotion.

Question 577

A RAG assistant is evaluated on only easy factual queries even though it also handles ambiguous refund requests. What test-data change improves coverage?

  1. Include representative conversation histories and turn-level expected outcomes.
  2. Use a separate held-out test set not consulted during prompt optimization.
  3. Add trustworthy ground-truth answers or expected outcomes to the evaluation dataset.
  4. Build a representative evaluation dataset including ordinary and ambiguous high-risk user intents.

Explanation

Correct Answer

D:

Test cases must reflect the variety of real requests; easy-only evaluation masks failures on important edge cases.

Incorrect Answers

A:

Include representative conversation histories and turn-level expected outcomes. It addresses another case: An assistant handles multi-turn conversations, but its benchmark tests only one isolated response at a time. For that case, the distinguishing justification or acceptance check is: Multi-turn evaluation can detect context handling errors that single-question datasets fail to reveal. The present case instead states: A RAG assistant is evaluated on only easy factual queries even though it also handles ambiguous refund requests. It calls for Build a representative evaluation dataset including ordinary and ambiguous high-risk user intents. Showing success for the other case would not establish: Test cases must reflect the variety of real requests; easy-only evaluation masks failures on important edge cases.

B:

The alternative, Use a separate held-out test set not consulted during prompt optimization, belongs to a different technical condition: A release candidate is judged on the same examples used repeatedly to tune its prompt. There the source-specific mechanism or proof is: An untouched holdout reduces the chance that a prompt is approved merely for memorizing known evaluation examples. The condition asked about here is: A RAG assistant is evaluated on only easy factual queries even though it also handles ambiguous refund requests. A suitable correction is Build a representative evaluation dataset including ordinary and ambiguous high-risk user intents; the targeted justification or result is: Test cases must reflect the variety of real requests; easy-only evaluation masks failures on important edge cases.

C:

Add trustworthy ground-truth answers or expected outcomes to the evaluation dataset. That operation would be relevant if: A support chatbot is being evaluated for exact answer correctness, but test data includes only prompts and model responses. Its corresponding reason or verification is: Reference answers enable task accuracy or completeness measurement beyond generic fluency scores on unlabelled outputs. However the problem in this question is: A RAG assistant is evaluated on only easy factual queries even though it also handles ambiguous refund requests. The requirement instead calls for Build a representative evaluation dataset including ordinary and ambiguous high-risk user intents; the question-specific check is: Test cases must reflect the variety of real requests; easy-only evaluation masks failures on important edge cases.

Question 578

A Foundry groundedness evaluator receives generated answers but no retrieved source context. Which dataset field should be included?

  1. Minimize and deidentify sensitive records while preserving the evaluation's necessary characteristics.
  2. Version the evaluation documents and record the exact policy snapshot used for scoring.
  3. Map the actual retrieved context to the evaluator's expected context input.
  4. Normalize dataset schema and configure evaluator column mapping to existing fields.

Explanation

Correct Answer

C:

Groundedness assessment requires supporting evidence to check whether a response follows retrieved documents rather than invented facts.

Incorrect Answers

A:

Minimize and deidentify sensitive records while preserving the evaluation's necessary characteristics. This is useful for the distinct issue: A test collection accidentally contains customer email addresses and account numbers that need not be present. The relevant technical rationale or evidence for that issue is: Test quality does not require uncontrolled personal identifiers; privacy safeguards should apply to evaluation corpora as well. Here the reported issue is: A Foundry groundedness evaluator receives generated answers but no retrieved source context. The relevant response is: Map the actual retrieved context to the evaluator's expected context input. The first result cannot verify this different correction: Groundedness assessment requires supporting evidence to check whether a response follows retrieved documents rather than invented facts.

B:

Version the evaluation documents and record the exact policy snapshot used for scoring. It addresses another case: A Foundry evaluation repeatedly uses stale customer policies despite new production document versions. For that case, the distinguishing justification or acceptance check is: Document provenance makes results comparable and prevents outdated reference content from masquerading as model failure. The present case instead states: A Foundry groundedness evaluator receives generated answers but no retrieved source context. It calls for Map the actual retrieved context to the evaluator's expected context input. Showing success for the other case would not establish: Groundedness assessment requires supporting evidence to check whether a response follows retrieved documents rather than invented facts.

D:

The alternative, Normalize dataset schema and configure evaluator column mapping to existing fields, belongs to a different technical condition: One evaluation run reports missing responses for several entries because source rows use different JSON property names. There the source-specific mechanism or proof is: Evaluation tooling requires consistent input names; mismatched fields can silently reduce sample coverage or fail outright. The condition asked about here is: A Foundry groundedness evaluator receives generated answers but no retrieved source context. A suitable correction is Map the actual retrieved context to the evaluator's expected context input; the targeted justification or result is: Groundedness assessment requires supporting evidence to check whether a response follows retrieved documents rather than invented facts.

Question 579

A support chatbot is being evaluated for exact answer correctness, but test data includes only prompts and model responses. What additional labels are needed?

  1. Include representative conversation histories and turn-level expected outcomes.
  2. Add trustworthy ground-truth answers or expected outcomes to the evaluation dataset.
  3. Map the actual retrieved context to the evaluator's expected context input.
  4. Minimize and deidentify sensitive records while preserving the evaluation's necessary characteristics.

Explanation

Correct Answer

B:

Reference answers enable task accuracy or completeness measurement beyond generic fluency scores on unlabelled outputs.

Incorrect Answers

A:

Include representative conversation histories and turn-level expected outcomes. A separate source case supports that approach when: An assistant handles multi-turn conversations, but its benchmark tests only one isolated response at a time. The supporting rationale or observable result is: Multi-turn evaluation can detect context handling errors that single-question datasets fail to reveal. This question describes a different issue: A support chatbot is being evaluated for exact answer correctness, but test data includes only prompts and model responses. Address it through Add trustworthy ground-truth answers or expected outcomes to the evaluation dataset, with the required justification or evidence: Reference answers enable task accuracy or completeness measurement beyond generic fluency scores on unlabelled outputs.

C:

Map the actual retrieved context to the evaluator's expected context input. This is useful for the distinct issue: A Foundry groundedness evaluator receives generated answers but no retrieved source context. The relevant technical rationale or evidence for that issue is: Groundedness assessment requires supporting evidence to check whether a response follows retrieved documents rather than invented facts. Here the reported issue is: A support chatbot is being evaluated for exact answer correctness, but test data includes only prompts and model responses. The relevant response is: Add trustworthy ground-truth answers or expected outcomes to the evaluation dataset. The first result cannot verify this different correction: Reference answers enable task accuracy or completeness measurement beyond generic fluency scores on unlabelled outputs.

D:

Minimize and deidentify sensitive records while preserving the evaluation's necessary characteristics. It addresses another case: A test collection accidentally contains customer email addresses and account numbers that need not be present. For that case, the distinguishing justification or acceptance check is: Test quality does not require uncontrolled personal identifiers; privacy safeguards should apply to evaluation corpora as well. The present case instead states: A support chatbot is being evaluated for exact answer correctness, but test data includes only prompts and model responses. It calls for Add trustworthy ground-truth answers or expected outcomes to the evaluation dataset. Showing success for the other case would not establish: Reference answers enable task accuracy or completeness measurement beyond generic fluency scores on unlabelled outputs.

Question 580

An evaluation dataset mixes tool-call traces, user messages, and generated answers into a single untyped column. What should be changed?

  1. Map each field to its documented query, response, context, and tool metadata role.
  2. Build a representative evaluation dataset including ordinary and ambiguous high-risk user intents.
  3. Version the evaluation documents and record the exact policy snapshot used for scoring.
  4. Use a separate held-out test set not consulted during prompt optimization.

Explanation

Correct Answer

A:

Evaluators need consistent typed inputs; ambiguous columns make scores misleading or cause validation failures.

Incorrect Answers

B:

Build a representative evaluation dataset including ordinary and ambiguous high-risk user intents. A separate source case supports that approach when: A RAG assistant is evaluated on only easy factual queries even though it also handles ambiguous refund requests. The supporting rationale or observable result is: Test cases must reflect the variety of real requests; easy-only evaluation masks failures on important edge cases. This question describes a different issue: An evaluation dataset mixes tool-call traces, user messages, and generated answers into a single untyped column. Address it through Map each field to its documented query, response, context, and tool metadata role, with the required justification or evidence: Evaluators need consistent typed inputs; ambiguous columns make scores misleading or cause validation failures.

C:

Version the evaluation documents and record the exact policy snapshot used for scoring. This is useful for the distinct issue: A Foundry evaluation repeatedly uses stale customer policies despite new production document versions. The relevant technical rationale or evidence for that issue is: Document provenance makes results comparable and prevents outdated reference content from masquerading as model failure. Here the reported issue is: An evaluation dataset mixes tool-call traces, user messages, and generated answers into a single untyped column. The relevant response is: Map each field to its documented query, response, context, and tool metadata role. The first result cannot verify this different correction: Evaluators need consistent typed inputs; ambiguous columns make scores misleading or cause validation failures.

D:

Use a separate held-out test set not consulted during prompt optimization. It addresses another case: A release candidate is judged on the same examples used repeatedly to tune its prompt. For that case, the distinguishing justification or acceptance check is: An untouched holdout reduces the chance that a prompt is approved merely for memorizing known evaluation examples. The present case instead states: An evaluation dataset mixes tool-call traces, user messages, and generated answers into a single untyped column. It calls for Map each field to its documented query, response, context, and tool metadata role. Showing success for the other case would not establish: Evaluators need consistent typed inputs; ambiguous columns make scores misleading or cause validation failures.

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