Practice Exams:

Microsoft AI-300 Dumps

EXAM: Operationalizing Machine Learning and Generative AI Solutions
Verified By Experts
Premium Bundle
$49.99

AI-300 Premium ETE File


  • Premium File 213 Questions & Answers
Whats Included:
  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates

Microsoft AI-300 Practice Test 17

Question 160

A university replaces a chain of emailed Python scripts with Azure ML pipeline components and needs each evaluation artifact traced to the correct training job. Which combined implementation and verification plan is the best match for this workload?

  1. Implement: Configure automated machine learning with time-aware or group-aware training and validation splits appropriate to the prediction task, and select metrics evaluated only on genuinely withheld examples. Verify: Run an AutoML experiment with the defined split policy, inspect train and validation identifiers and test the selected candidate on a separate future-dated or unseen-group holdout.
  2. Implement: Orchestrate feature preparation, validation, training and evaluation as pipeline components connected through versioned typed inputs and outputs, adding failure gates before model registration. Verify: Execute two controlled runs with different hyperparameters and query experiment history to confirm both runs show their distinct metrics, inputs and output artifacts.
  3. Implement: Orchestrate feature preparation, validation, training and evaluation as pipeline components connected through versioned typed inputs and outputs, adding failure gates before model registration. Verify: Submit a disposable pipeline with a changed input asset, confirm downstream steps consume that run's output and verify a failed validation step prevents the promotion component from executing.
  4. Implement: Log each training attempt to MLflow with its parameter values, dataset and environment versions, selected evaluation metrics and a distinct run identifier under the correct experiment. Verify: Submit a disposable pipeline with a changed input asset, confirm downstream steps consume that run's output and verify a failed validation step prevents the promotion component from executing.

Answer & Explanations

Correct answer: C

A.

This action/test pair addresses “Prevent training validation leakage during Azure automated ML experiments”. Configure automated machine learning with time-aware or group-aware training and validation splits appropriate to the prediction task, and select metrics evaluated only on genuinely withheld examples. Run an AutoML experiment with the defined split policy, inspect train and validation identifiers and test the selected candidate on a separate future-dated or unseen-group holdout. The case instead calls for: Orchestrate feature preparation, validation, training and evaluation as pipeline components connected through versioned typed inputs and outputs, adding failure gates before model registration.

B.

The action is applicable: Orchestrate feature preparation, validation, training and evaluation as pipeline components connected through versioned typed inputs and outputs, adding failure gates before model registration. But its proposed test—Execute two controlled runs with different hyperparameters and query experiment history to confirm both runs show their distinct metrics, inputs and output artifacts.—addresses “Track Azure Machine Learning experiments and metrics with MLflow run identity”. This requirement is confirmed instead by: Submit a disposable pipeline with a changed input asset, confirm downstream steps consume that run's output and verify a failed validation step prevents the promotion component from executing.

C.

The question is set in this situation: A university replaces a chain of emailed Python scripts with Azure ML pipeline components and needs each evaluation artifact traced to the correct training job. Orchestrate feature preparation, validation, training and evaluation as pipeline components connected through versioned typed inputs and outputs, adding failure gates before model registration. Confirm it through this test: Submit a disposable pipeline with a changed input asset, confirm downstream steps consume that run's output and verify a failed validation step prevents the promotion component from executing. Managed pipeline dependency graphs schedule steps according to declared data dependencies, while launching shell commands in an arbitrary order risks evaluating a stale previous-run model artifact.

D.

The test would be relevant: Submit a disposable pipeline with a changed input asset, confirm downstream steps consume that run's output and verify a failed validation step prevents the promotion component from executing. But the offered implementation—Log each training attempt to MLflow with its parameter values, dataset and environment versions, selected evaluation metrics and a distinct run identifier under the correct experiment.—addresses “Track Azure Machine Learning experiments and metrics with MLflow run identity”. The action needed here is: Orchestrate feature preparation, validation, training and evaluation as pipeline components connected through versioned typed inputs and outputs, adding failure gates before model registration.

Question 161

A lending platform trained with seven-day transaction features and needs online inference to recover the same definitions instead of raw totals from an unrelated API. Which Azure AI operations action best addresses this requirement?

  1. Create a parallel named Azure Machine Learning online deployment for the candidate model, send a limited test or canary traffic share, monitor guardrail metrics and gradually shift routing after approval.
  2. Record the registered model version, approval decision, deployment associations and retirement state so unused models can be archived while active or rollback versions remain discoverable.
  3. Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.
  4. Before release, assess model errors and harmful disparities across relevant permitted cohorts using slice-level performance, explainability and reliability checks, with documented mitigations and owners.

Answer & Explanations

Correct answer: C

A.

Create a parallel named Azure Machine Learning online deployment for the candidate model, send a limited test or canary traffic share, monitor guardrail metrics and gradually shift routing after approval. This addresses “Apply blue-green or canary rollout with explicit endpoint traffic weights”, where the governing principle is: Multiple deployments behind one endpoint can receive configured traffic portions, enabling controlled promotion while the stable deployment remains a rollback target instead of being overwritten. The stated case instead calls for: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

B.

Record the registered model version, approval decision, deployment associations and retirement state so unused models can be archived while active or rollback versions remain discoverable. This addresses “Archive superseded model versions without losing active deployment lineage”, where the governing principle is: Model registry lifecycle and endpoint configuration are distinct: archiving a model should not silently replace a running production deployment, and deleting rollback artifacts can break recovery. The stated case instead calls for: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

C.

Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew. Verify the result with this test: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

D.

Before release, assess model errors and harmful disparities across relevant permitted cohorts using slice-level performance, explainability and reliability checks, with documented mitigations and owners. This addresses “Evaluate responsible AI performance across representative data slices”, where the governing principle is: A single aggregate accuracy or AUC value can hide materially different false-negative rates across groups; approval requires a task-relevant evaluation beyond overall score. The stated case instead calls for: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

Question 162

A lending platform trained with seven-day transaction features and needs online inference to recover the same definitions instead of raw totals from an unrelated API. The engineer proposes the following change: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. Which mechanism explains why it works?

  1. A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.
  2. Model registry lifecycle and endpoint configuration are distinct: archiving a model should not silently replace a running production deployment, and deleting rollback artifacts can break recovery.
  3. Multiple deployments behind one endpoint can receive configured traffic portions, enabling controlled promotion while the stable deployment remains a rollback target instead of being overwritten.
  4. A single aggregate accuracy or AUC value can hide materially different false-negative rates across groups; approval requires a task-relevant evaluation beyond overall score.

Answer & Explanations

Correct answer: A

A.

A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew. The corresponding implementation is: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

B.

Model registry lifecycle and endpoint configuration are distinct: archiving a model should not silently replace a running production deployment, and deleting rollback artifacts can break recovery. That explains “Archive superseded model versions without losing active deployment lineage”, not the mechanism behind the described change. The relevant principle is: A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

C.

Multiple deployments behind one endpoint can receive configured traffic portions, enabling controlled promotion while the stable deployment remains a rollback target instead of being overwritten. That explains “Apply blue-green or canary rollout with explicit endpoint traffic weights”, not the mechanism behind the described change. The relevant principle is: A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

D.

A single aggregate accuracy or AUC value can hide materially different false-negative rates across groups; approval requires a task-relevant evaluation beyond overall score. That explains “Evaluate responsible AI performance across representative data slices”, not the mechanism behind the described change. The relevant principle is: A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

Question 163

A lending platform trained with seven-day transaction features and needs online inference to recover the same definitions instead of raw totals from an unrelated API. After implementing this change—Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.—which test most directly validates its intended outcome?

  1. Trace the deployed endpoint to its immutable registered model version, archive a test unused version and confirm current inference and an authorized rollback candidate remain available.
  2. Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.
  3. Produce separate confusion matrices and calibrated error metrics for authorized cohorts, document known limitations and verify remediation improves slices without an unacceptable global regression.
  4. Confirm requests split according to the chosen traffic settings, compare errors and latency by deployment, then perform an intentional rollback to the old model without rebuilding its container.

Answer & Explanations

Correct answer: B

A.

Trace the deployed endpoint to its immutable registered model version, archive a test unused version and confirm current inference and an authorized rollback candidate remain available. That would help test “Archive superseded model versions without losing active deployment lineage”, not the outcome required in this scenario. The relevant acceptance check is: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

B.

Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema. This check is decisive because a model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

C.

Produce separate confusion matrices and calibrated error metrics for authorized cohorts, document known limitations and verify remediation improves slices without an unacceptable global regression. That would help test “Evaluate responsible AI performance across representative data slices”, not the outcome required in this scenario. The relevant acceptance check is: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

D.

Confirm requests split according to the chosen traffic settings, compare errors and latency by deployment, then perform an intentional rollback to the old model without rebuilding its container. That would help test “Apply blue-green or canary rollout with explicit endpoint traffic weights”, not the outcome required in this scenario. The relevant acceptance check is: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

Question 164

A lending platform trained with seven-day transaction features and needs online inference to recover the same definitions instead of raw totals from an unrelated API. Which proposed shortcut most directly risks incorrect predictions after production feature retrieval diverges from training assumptions and should therefore be rejected?

  1. Approve a credit model solely from high overall AUC and assume no subgroup fairness or reliability assessment is necessary because it is deployed in Azure.
  2. Delete every prior model version immediately after a new score improves and assume endpoint rollback will recreate historical model files from memory.
  3. Overwrite the only production scoring deployment in place and assume a model registry version string alone provides instant traffic rollback.
  4. Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields.

Answer & Explanations

Correct answer: D

A.

Approve a credit model solely from high overall AUC and assume no subgroup fairness or reliability assessment is necessary because it is deployed in Azure. That shortcut risks unnoticed unequal error burdens and unreliable decisions under a misleading aggregate metric. Here the specific shortcut is: Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields.

B.

Delete every prior model version immediately after a new score improves and assume endpoint rollback will recreate historical model files from memory. That shortcut risks irreversible loss of rollback artifacts after a poor model promotion. Here the specific shortcut is: Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields.

C.

Overwrite the only production scoring deployment in place and assume a model registry version string alone provides instant traffic rollback. That shortcut risks unrecoverable live prediction outages during an untested model rollout. Here the specific shortcut is: Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields.

D.

Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields. The foreseeable failure is incorrect predictions after production feature retrieval diverges from training assumptions. The safer response is: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

Question 165

A lending platform trained with seven-day transaction features and needs online inference to recover the same definitions instead of raw totals from an unrelated API. Which combined implementation and verification plan is the best match for this workload?

  1. Implement: Before release, assess model errors and harmful disparities across relevant permitted cohorts using slice-level performance, explainability and reliability checks, with documented mitigations and owners. Verify: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.
  2. Implement: Record the registered model version, approval decision, deployment associations and retirement state so unused models can be archived while active or rollback versions remain discoverable. Verify: Trace the deployed endpoint to its immutable registered model version, archive a test unused version and confirm current inference and an authorized rollback candidate remain available.
  3. Implement: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. Verify: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.
  4. Implement: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. Verify: Produce separate confusion matrices and calibrated error metrics for authorized cohorts, document known limitations and verify remediation improves slices without an unacceptable global regression.

Answer & Explanations

Correct answer: C

A.

The test would be relevant: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema. But the offered implementation—Before release, assess model errors and harmful disparities across relevant permitted cohorts using slice-level performance, explainability and reliability checks, with documented mitigations and owners.—addresses “Evaluate responsible AI performance across representative data slices”. The action needed here is: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

B.

This action/test pair addresses “Archive superseded model versions without losing active deployment lineage”. Record the registered model version, approval decision, deployment associations and retirement state so unused models can be archived while active or rollback versions remain discoverable. Trace the deployed endpoint to its immutable registered model version, archive a test unused version and confirm current inference and an authorized rollback candidate remain available. The case instead calls for: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

C.

Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. Confirm it through this test: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema. A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

D.

The action is applicable: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. But its proposed test—Produce separate confusion matrices and calibrated error metrics for authorized cohorts, document known limitations and verify remediation improves slices without an unacceptable global regression.—addresses “Evaluate responsible AI performance across representative data slices”. This requirement is confirmed instead by: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

Question 166

A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Which Azure AI operations action best addresses this requirement?

  1. Before release, assess model errors and harmful disparities across relevant permitted cohorts using slice-level performance, explainability and reliability checks, with documented mitigations and owners.
  2. Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.
  3. Record the registered model version, approval decision, deployment associations and retirement state so unused models can be archived while active or rollback versions remain discoverable.
  4. Monitor managed inference request count, throttling, error rates, latency percentiles and compute saturation with scoped alerts and on-call routing, separately tracking prediction quality.

Answer & Explanations

Correct answer: B

A.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Before release, assess model errors and harmful disparities across relevant permitted cohorts using slice-level performance, explainability and reliability checks, with documented mitigations and owners. This addresses “Evaluate responsible AI performance across representative data slices”, where the governing principle is: A single aggregate accuracy or AUC value can hide materially different false-negative rates across groups; approval requires a task-relevant evaluation beyond overall score. The stated case instead calls for: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

B.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew. Verify the result with this test: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

C.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Record the registered model version, approval decision, deployment associations and retirement state so unused models can be archived while active or rollback versions remain discoverable. This addresses “Archive superseded model versions without losing active deployment lineage”, where the governing principle is: Model registry lifecycle and endpoint configuration are distinct: archiving a model should not silently replace a running production deployment, and deleting rollback artifacts can break recovery. The stated case instead calls for: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

D.

Monitor managed inference request count, throttling, error rates, latency percentiles and compute saturation with scoped alerts and on-call routing, separately tracking prediction quality. This addresses “Alert on production model serving health independently from accuracy drift”, where the governing principle is: Serving availability describes the operational request path while model quality describes prediction usefulness; an endpoint can return fast HTTP 200 responses with a drifting algorithm. The stated case instead calls for: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

Question 167

A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. The engineer proposes the following change: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. Which mechanism explains why it works?

  1. A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.
  2. Model registry lifecycle and endpoint configuration are distinct: archiving a model should not silently replace a running production deployment, and deleting rollback artifacts can break recovery.
  3. Serving availability describes the operational request path while model quality describes prediction usefulness; an endpoint can return fast HTTP 200 responses with a drifting algorithm.
  4. A single aggregate accuracy or AUC value can hide materially different false-negative rates across groups; approval requires a task-relevant evaluation beyond overall score.

Answer & Explanations

Correct answer: A

A.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew. The corresponding implementation is: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

B.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Model registry lifecycle and endpoint configuration are distinct: archiving a model should not silently replace a running production deployment, and deleting rollback artifacts can break recovery. That explains “Archive superseded model versions without losing active deployment lineage”, not the mechanism behind the described change. The relevant principle is: A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

C.

Serving availability describes the operational request path while model quality describes prediction usefulness; an endpoint can return fast HTTP 200 responses with a drifting algorithm. That explains “Alert on production model serving health independently from accuracy drift”, not the mechanism behind the described change. The relevant principle is: A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

D.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. A single aggregate accuracy or AUC value can hide materially different false-negative rates across groups; approval requires a task-relevant evaluation beyond overall score. That explains “Evaluate responsible AI performance across representative data slices”, not the mechanism behind the described change. The relevant principle is: A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

Question 168

A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. After implementing this change—Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.—which test most directly validates its intended outcome?

  1. Produce separate confusion matrices and calibrated error metrics for authorized cohorts, document known limitations and verify remediation improves slices without an unacceptable global regression.
  2. Trace the deployed endpoint to its immutable registered model version, archive a test unused version and confirm current inference and an authorized rollback candidate remain available.
  3. Generate a harmless burst of test requests, inspect p95 latency and throttling in monitoring, confirm an alert reaches its owner and leave model-quality tracking as a separate gate.
  4. Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

Answer & Explanations

Correct answer: D

A.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Produce separate confusion matrices and calibrated error metrics for authorized cohorts, document known limitations and verify remediation improves slices without an unacceptable global regression. That would help test “Evaluate responsible AI performance across representative data slices”, not the outcome required in this scenario. The relevant acceptance check is: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

B.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Trace the deployed endpoint to its immutable registered model version, archive a test unused version and confirm current inference and an authorized rollback candidate remain available. That would help test “Archive superseded model versions without losing active deployment lineage”, not the outcome required in this scenario. The relevant acceptance check is: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

C.

Generate a harmless burst of test requests, inspect p95 latency and throttling in monitoring, confirm an alert reaches its owner and leave model-quality tracking as a separate gate. That would help test “Alert on production model serving health independently from accuracy drift”, not the outcome required in this scenario. The relevant acceptance check is: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

D.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema. This check is decisive because a model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

Question 169

A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Which proposed shortcut most directly risks incorrect predictions after production feature retrieval diverges from training assumptions and should therefore be rejected?

  1. Delete every prior model version immediately after a new score improves and assume endpoint rollback will recreate historical model files from memory.
  2. Approve a credit model solely from high overall AUC and assume no subgroup fairness or reliability assessment is necessary because it is deployed in Azure.
  3. Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields.
  4. Show a green model-accuracy dashboard and assume it replaces alerts for endpoint 503 errors, high latency or exhausted compute quotas.

Answer & Explanations

Correct answer: C

A.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Delete every prior model version immediately after a new score improves and assume endpoint rollback will recreate historical model files from memory. That shortcut risks irreversible loss of rollback artifacts after a poor model promotion. Here the specific shortcut is: Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields.

B.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Approve a credit model solely from high overall AUC and assume no subgroup fairness or reliability assessment is necessary because it is deployed in Azure. That shortcut risks unnoticed unequal error burdens and unreliable decisions under a misleading aggregate metric. Here the specific shortcut is: Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields.

C.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields. The foreseeable failure is incorrect predictions after production feature retrieval diverges from training assumptions. The safer response is: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

D.

Show a green model-accuracy dashboard and assume it replaces alerts for endpoint 503 errors, high latency or exhausted compute quotas. That shortcut risks undetected inference service outages despite apparently acceptable offline validation scores. Here the specific shortcut is: Deploy a trained risk model without its feature lookup or transformation contract and let each service invent independently named input fields.

Question 170

A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Which combined implementation and verification plan is the best match for this workload?

  1. Implement: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. Verify: Produce separate confusion matrices and calibrated error metrics for authorized cohorts, document known limitations and verify remediation improves slices without an unacceptable global regression.
  2. Implement: Record the registered model version, approval decision, deployment associations and retirement state so unused models can be archived while active or rollback versions remain discoverable. Verify: Trace the deployed endpoint to its immutable registered model version, archive a test unused version and confirm current inference and an authorized rollback candidate remain available.
  3. Implement: Before release, assess model errors and harmful disparities across relevant permitted cohorts using slice-level performance, explainability and reliability checks, with documented mitigations and owners. Verify: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.
  4. Implement: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. Verify: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

Answer & Explanations

Correct answer: D

A.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. The action is applicable: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. But its proposed test—Produce separate confusion matrices and calibrated error metrics for authorized cohorts, document known limitations and verify remediation improves slices without an unacceptable global regression.—addresses “Evaluate responsible AI performance across representative data slices”. This requirement is confirmed instead by: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema.

B.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. This action/test pair addresses “Archive superseded model versions without losing active deployment lineage”. Record the registered model version, approval decision, deployment associations and retirement state so unused models can be archived while active or rollback versions remain discoverable. Trace the deployed endpoint to its immutable registered model version, archive a test unused version and confirm current inference and an authorized rollback candidate remain available. The case instead calls for: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

C.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. The test would be relevant: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema. But the offered implementation—Before release, assess model errors and harmful disparities across relevant permitted cohorts using slice-level performance, explainability and reliability checks, with documented mitigations and owners.—addresses “Evaluate responsible AI performance across representative data slices”. The action needed here is: Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training.

D.

The question is set in this situation: A manufacturing predictor expects sensor aggregates from an approved window and must prevent a serving service from sending instant readings that were never used during training. Register the feature lookup configuration, expected feature names, entity keys and point-in-time behavior alongside the model version so inference uses the same feature definitions approved in training. Confirm it through this test: Run a shadow inference test for known entities, compare retrieved feature fields and timestamps with the registered specification and confirm model inputs match the validated training schema. A model binary alone does not guarantee online predictions retrieve features with matching schema or freshness; mismatched serving transformations cause silent training-serving skew.

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