Regulators require that each declined loan application come with the per-feature contributions that drove the score, computed at the moment of the request. Which capability provides this?
Show answer and explanation
Answer: D
Vertex AI Explainable AI returns per-feature attributions with each online prediction when explanations are enabled on the endpoint, using methods such as integrated gradients or sampled Shapley. That gives the per-decision contribution breakdown the regulator is asking for, generated at inference time. Audit logs record who made a call, trace spans record timing, and a hypothesis test on training data is a population-level statistic, so none of them explain an individual score.
Why the other options are wrong
- A. Audit logs prove that a call happened and by whom, but they do not contain the feature-level reasons behind the score.
- B. Trace spans measure where time goes during a request, which is useful for performance but silent on why the model scored the applicant as it did.
- C. A hypothesis test on the training table describes the population used to fit the model, not the contribution of each feature for one declined applicant.