A. Incorrect.
Evaluate only on one prompt written by a developer is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Consistent evaluation prevents anecdotal model selection.

B. Incorrect.
Use cluster CPU utilization as the quality metric is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Consistent evaluation prevents anecdotal model selection.

C. Correct.
Consistent evaluation prevents anecdotal model selection.

D. Incorrect.
Choose the largest model without testing is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Consistent evaluation prevents anecdotal model selection.

A. Incorrect.
Keep a driver-only cluster idle is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Job compute starts for the run and terminates after completion.

B. Correct.
Job compute starts for the run and terminates after completion.

C. Incorrect.
Use a permanently running all-purpose GPU cluster is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Job compute starts for the run and terminates after completion.

D. Incorrect.
Use a SQL warehouse for PyTorch training is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Job compute starts for the run and terminates after completion.

A. Correct.
Auditability requires tying the artifact to exact data and code context.

B. Incorrect.
Only the author's first name is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Auditability requires tying the artifact to exact data and code context.

C. Incorrect.
Only the cluster color theme is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Auditability requires tying the artifact to exact data and code context.

D. Incorrect.
Sorting descending returns the largest region_id values first, while the requirement asks for the smallest three.

A. Incorrect.
Whether the notebook has too many cells is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Leakage occurs when training uses future or target-derived information.

B. Incorrect.
Whether the cluster has autoscaling disabled is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Leakage occurs when training uses future or target-derived information.

C. Incorrect.
Whether the table is OPTIMIZEd is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Leakage occurs when training uses future or target-derived information.

D. Correct.
Leakage occurs when training uses future or target-derived information.

A. Incorrect.
Manually paste predictions into a dashboard is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Batch inference is efficient for large non-real-time scoring workloads.

B. Incorrect.
Store predictions in notebook output only is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Batch inference is efficient for large non-real-time scoring workloads.

C. Correct.
Batch inference is efficient for large non-real-time scoring workloads.

D. Incorrect.
Call a REST endpoint once per row from a laptop is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Batch inference is efficient for large non-real-time scoring workloads.

A. Incorrect.
Automatically approve every prediction is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Human-in-the-loop review is useful for risk-sensitive decisions.

B. Correct.
Human-in-the-loop review is useful for risk-sensitive decisions.

C. Incorrect.
Hide confidence scores is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Human-in-the-loop review is useful for risk-sensitive decisions.

D. Incorrect.
Disable logging is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Human-in-the-loop review is useful for risk-sensitive decisions.

A. Correct.
Reproducibility depends on the full execution environment.

B. Incorrect.
Only dashboard names is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Reproducibility depends on the full execution environment.

C. Incorrect.
Only notebook cell order is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Reproducibility depends on the full execution environment.

D. Incorrect.
Only user display names is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Reproducibility depends on the full execution environment.

A. Incorrect.
Guaranteed lower latency is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Same-sample evaluation isolates model behavior from data differences.

B. Incorrect.
No need for metrics is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Same-sample evaluation isolates model behavior from data differences.

C. Incorrect.
Automatic data deletion is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Same-sample evaluation isolates model behavior from data differences.

D. Correct.
Same-sample evaluation isolates model behavior from data differences.

A. Incorrect.
Browser bookmarks is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Privileges control who can read or modify governed assets.

B. Incorrect.
Increasing shuffle partitions is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Privileges control who can read or modify governed assets.

C. Correct.
Privileges control who can read or modify governed assets.

D. Incorrect.
Notebook output clearing is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Privileges control who can read or modify governed assets.

A. Incorrect.
The workspace trash folder is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Serving observability exposes endpoint health and performance.

B. Correct.
Serving observability exposes endpoint health and performance.

C. Incorrect.
The table schema only is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Serving observability exposes endpoint health and performance.

D. Incorrect.
A cluster library list only is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Serving observability exposes endpoint health and performance.