Microsoft AI-300 (Operationalizing Machine Learning and Generative AI Solutions) Practice Test 22

Question 210

A model depends on consistent feature retrieval during training and inference. An operator is tasked to specify safe behavior when an online feature source is temporarily unavailable. Which mechanism provides that capability?

  1. Online feature freshness
  2. Feature retrieval specification
  3. Feature schema version
  4. Point-in-time feature joins
  5. Missing-feature fallback

Explanation

Correct Answer

E:

A fallback policy defines whether to use an approved default, cached value, or fail closed. It should preserve correctness and risk requirements rather than invent arbitrary values.

Incorrect Answers

A:

Freshness monitoring detects delayed or stale feature updates; it is not the same as detecting long-term distribution drift across populations; the outstanding need is to specify safe behavior when an online feature source is temporarily unavailable.

B:

A feature retrieval contract makes dependencies explicit for serving and reproducibility; a bare weights file is insufficient when inference requires external features; this approach leaves the team unable to specify safe behavior when an online feature source is temporarily unavailable.

C:

Feature schema versioning helps avoid incompatible inference payloads; silently reordering inputs can corrupt predictions without an obvious application error; the needed capability is to specify safe behavior when an online feature source is temporarily unavailable.

D:

Point-in-time joins respect event timestamps and feature availability; ordinary latest-value joins can leak later information into historical examples; the requirement, by contrast, is to specify safe behavior when an online feature source is temporarily unavailable.

Question 211

Governance separates experimental model artifacts from approved production releases. The immediate design goal is to create a model asset that points to a specific training output for subsequent deployment. Which approach addresses that goal?

  1. Model rollback reference
  2. Model registration from run
  3. Staging model version
  4. Model archive state

Explanation

Correct Answer

B:

Registration supplies an identifiable artifact and version, with possible lineage to the producing run. It does not itself allocate a live serving endpoint.

Incorrect Answers

A:

A pinned previous model version provides a reproducible rollback target; re-training from scratch during an outage would add unnecessary uncertainty and delay; the requirement, by contrast, is to create a model asset that points to a specific training output for subsequent deployment.

C:

Staging validates application contracts and runtime behavior; it cannot be replaced by a good offline metric alone; the specified outcome instead requires the team to create a model asset that points to a specific training output for subsequent deployment.

D:

Archiving is a lifecycle governance action that can reduce confusion about supported releases; it differs from deleting evidence of earlier deployments; the implementation under review needs to create a model asset that points to a specific training output for subsequent deployment.

Question 212

Governance separates experimental model artifacts from approved production releases. The team is required to require an explicit review of quality and risk results before switching serving traffic to a candidate. Which option is the best fit?

  1. Production approval gate
  2. Release provenance manifest
  3. Model archive state
  4. Model rollback reference

Explanation

Correct Answer

A:

Approval separates technical training completion from authorization to release. Without a gate, a merely successful job could promote a harmful model.

Incorrect Answers

B:

A release manifest links the components that produced and qualified the model; merely recording its display name is too weak for incident reconstruction; the requested task remains to require an explicit review of quality and risk results before switching serving traffic to a candidate.

C:

Archiving is a lifecycle governance action that can reduce confusion about supported releases; it differs from deleting evidence of earlier deployments; this alternative does not achieve the requirement to require an explicit review of quality and risk results before switching serving traffic to a candidate.

D:

A pinned previous model version provides a reproducible rollback target; re-training from scratch during an outage would add unnecessary uncertainty and delay; the actual design criterion is to require an explicit review of quality and risk results before switching serving traffic to a candidate.

Question 213

Governance separates experimental model artifacts from approved production releases. An operator is tasked to deploy a candidate to a nonproduction location for integration and load testing before production. Which mechanism provides that capability?

  1. Production approval gate
  2. Model registration from run
  3. Release provenance manifest
  4. Staging model version

Explanation

Correct Answer

D:

Staging validates application contracts and runtime behavior. It cannot be replaced by a good offline metric alone.

Incorrect Answers

A:

Approval separates technical training completion from authorization to release; without a gate, a merely successful job could promote a harmful model; the actual design criterion is to deploy a candidate to a nonproduction location for integration and load testing before production.

B:

Registration supplies an identifiable artifact and version, with possible lineage to the producing run; it does not itself allocate a live serving endpoint; the outstanding need is to deploy a candidate to a nonproduction location for integration and load testing before production.

C:

A release manifest links the components that produced and qualified the model; merely recording its display name is too weak for incident reconstruction; this approach leaves the team unable to deploy a candidate to a nonproduction location for integration and load testing before production.

Question 214

Governance separates experimental model artifacts from approved production releases. The immediate design goal is to retain historic versions for audit while stopping them from appearing as actively supported candidates. Which approach addresses that goal?

  1. Model archive state
  2. Production approval gate
  3. Staging model version
  4. Model registration from run

Explanation

Correct Answer

A:

Archiving is a lifecycle governance action that can reduce confusion about supported releases. It differs from deleting evidence of earlier deployments.

Incorrect Answers

B:

Approval separates technical training completion from authorization to release; without a gate, a merely successful job could promote a harmful model; the requirement, by contrast, is to retain historic versions for audit while stopping them from appearing as actively supported candidates.

C:

Staging validates application contracts and runtime behavior; it cannot be replaced by a good offline metric alone; the team would still need to retain historic versions for audit while stopping them from appearing as actively supported candidates.

D:

Registration supplies an identifiable artifact and version, with possible lineage to the producing run; it does not itself allocate a live serving endpoint; the specified outcome instead requires the team to retain historic versions for audit while stopping them from appearing as actively supported candidates.

Question 215

Governance separates experimental model artifacts from approved production releases. The team is required to keep a known-good, immutable artifact version available for quick recovery from a failed release. Which option is the best fit?

  1. Staging model version
  2. Production approval gate
  3. Model rollback reference
  4. Model registration from run

Explanation

Correct Answer

C:

A pinned previous model version provides a reproducible rollback target. Re-training from scratch during an outage would add unnecessary uncertainty and delay.

Incorrect Answers

A:

Staging validates application contracts and runtime behavior; it cannot be replaced by a good offline metric alone; the specified outcome instead requires the team to keep a known-good, immutable artifact version available for quick recovery from a failed release.

B:

Approval separates technical training completion from authorization to release; without a gate, a merely successful job could promote a harmful model; the implementation under review needs to keep a known-good, immutable artifact version available for quick recovery from a failed release.

D:

Registration supplies an identifiable artifact and version, with possible lineage to the producing run; it does not itself allocate a live serving endpoint; this alternative does not achieve the requirement to keep a known-good, immutable artifact version available for quick recovery from a failed release.

Question 216

Governance separates experimental model artifacts from approved production releases. An operator is tasked to record code, data, environment, model version, and test evidence for an approved deployment. Which mechanism provides that capability?

  1. Model archive state
  2. Release provenance manifest
  3. Model rollback reference
  4. Staging model version

Explanation

Correct Answer

B:

A release manifest links the components that produced and qualified the model. Merely recording its display name is too weak for incident reconstruction.

Incorrect Answers

A:

Archiving is a lifecycle governance action that can reduce confusion about supported releases; it differs from deleting evidence of earlier deployments; this alternative does not achieve the requirement to record code, data, environment, model version, and test evidence for an approved deployment.

C:

A pinned previous model version provides a reproducible rollback target; re-training from scratch during an outage would add unnecessary uncertainty and delay; the outstanding need is to record code, data, environment, model version, and test evidence for an approved deployment.

D:

Staging validates application contracts and runtime behavior; it cannot be replaced by a good offline metric alone; this approach leaves the team unable to record code, data, environment, model version, and test evidence for an approved deployment.

Question 217

An AI risk review evaluates whether a predictive model treats important populations consistently. The immediate design goal is to compare model error rates across defined demographic or operational subgroups rather than only global accuracy. Which approach addresses that goal?

  1. Data representation audit
  2. Performance parity assessment
  3. Protected-feature leakage check
  4. Group fairness analysis
  5. Fairness mitigation experiment

Explanation

Correct Answer

D:

Fairness analysis tests distributional performance differences across groups. It requires suitable labels, context, and careful interpretation of chosen fairness criteria.

Incorrect Answers

A:

Representation checks examine whether available training data covers intended users; post-training threshold changes alone cannot repair missing evidence for a population; this approach leaves the team unable to compare model error rates across defined demographic or operational subgroups rather than only global accuracy.

B:

Segment-level parity analysis reveals uneven error distributions; the aggregate accuracy metric may conceal that disparity; the needed capability is to compare model error rates across defined demographic or operational subgroups rather than only global accuracy.

C:

Proxy analysis investigates indirect pathways to disparate outcomes; dropping the protected attribute does not guarantee fair treatment; the requirement, by contrast, is to compare model error rates across defined demographic or operational subgroups rather than only global accuracy.

E:

Mitigation can alter tradeoffs between group metrics and other performance dimensions; it should be evaluated empirically rather than assumed harmless; the specified outcome instead requires the team to compare model error rates across defined demographic or operational subgroups rather than only global accuracy.

Question 218

An AI risk review evaluates whether a predictive model treats important populations consistently. The team is required to identify underrepresented populations before training and plan representative collection or weighting. Which option is the best fit?

  1. Fairness mitigation experiment
  2. Data representation audit
  3. Group fairness analysis
  4. Fairness documentation

Explanation

Correct Answer

B:

Representation checks examine whether available training data covers intended users. Post-training threshold changes alone cannot repair missing evidence for a population.

Incorrect Answers

A:

Mitigation can alter tradeoffs between group metrics and other performance dimensions; it should be evaluated empirically rather than assumed harmless; the team would still need to identify underrepresented populations before training and plan representative collection or weighting.

C:

Fairness analysis tests distributional performance differences across groups; it requires suitable labels, context, and careful interpretation of chosen fairness criteria; the implementation under review needs to identify underrepresented populations before training and plan representative collection or weighting.

D:

Documented results make assumptions and residual risks reviewable; a one-number model score cannot describe all fairness properties; the requested task remains to identify underrepresented populations before training and plan representative collection or weighting.

Question 219

An AI risk review evaluates whether a predictive model treats important populations consistently. An operator is tasked to investigate why recall is much worse for one customer segment despite strong overall performance. Which mechanism provides that capability?

  1. Group fairness analysis
  2. Performance parity assessment
  3. Data representation audit
  4. Fairness documentation

Explanation

Correct Answer

B:

Segment-level parity analysis reveals uneven error distributions. The aggregate accuracy metric may conceal that disparity.

Incorrect Answers

A:

Fairness analysis tests distributional performance differences across groups; it requires suitable labels, context, and careful interpretation of chosen fairness criteria; the requested task remains to investigate why recall is much worse for one customer segment despite strong overall performance.

C:

Representation checks examine whether available training data covers intended users; post-training threshold changes alone cannot repair missing evidence for a population; the actual design criterion is to investigate why recall is much worse for one customer segment despite strong overall performance.

D:

Documented results make assumptions and residual risks reviewable; a one-number model score cannot describe all fairness properties; the outstanding need is to investigate why recall is much worse for one customer segment despite strong overall performance.

Question 220

An AI risk review evaluates whether a predictive model treats important populations consistently. The immediate design goal is to compare an approved mitigation with the baseline under both equity and predictive performance metrics. Which approach addresses that goal?

  1. Fairness mitigation experiment
  2. Protected-feature leakage check
  3. Fairness documentation
  4. Data representation audit

Explanation

Correct Answer

A:

Mitigation can alter tradeoffs between group metrics and other performance dimensions. It should be evaluated empirically rather than assumed harmless.

Incorrect Answers

B:

Proxy analysis investigates indirect pathways to disparate outcomes; dropping the protected attribute does not guarantee fair treatment; this approach leaves the team unable to compare an approved mitigation with the baseline under both equity and predictive performance metrics.

C:

Documented results make assumptions and residual risks reviewable; a one-number model score cannot describe all fairness properties; the needed capability is to compare an approved mitigation with the baseline under both equity and predictive performance metrics.

D:

Representation checks examine whether available training data covers intended users; post-training threshold changes alone cannot repair missing evidence for a population; the requirement, by contrast, is to compare an approved mitigation with the baseline under both equity and predictive performance metrics.

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