Question 620
At an executive oversight process for an AI service, an AI project leader faces this situation: An ML result cannot be replicated because random seeds, dependencies and feature definitions were not recorded. Responsible AI controls need measurable evidence and assigned decision owners. What result should the project owner require before accepting the change?
Explanation
Correct Answer
In an executive oversight process for an AI service, the decisive check is: An authorized repeat experiment reproduces the expected metrics within documented tolerances. That shows the planned action has worked: Capture configuration, environment, seed, code commit and data versions for each training run. The governing reason is: Reproducibility requires sufficient experiment metadata to reconstruct a training configuration and inputs. A separate successful activity would not establish this specific requirement.
Incorrect Answers
Model-serving endpoints can expose training-derived or user-specific information if access is uncontrolled. This alternative mechanism does not match the current fault.
Stakeholders need traceable descriptions of data sources and transformations to assess output limitations. Fixing that different mechanism would not address the current root cause: reproducibility requires sufficient experiment metadata to reconstruct a training configuration and inputs.
Unbalanced coverage of affected populations can produce systematic differences in model quality. Its verification is distinct from the reported condition: An ML result cannot be replicated because random seeds, dependencies and feature definitions were not recorded. Responsible AI controls need measurable evidence and assigned decision owners.