By 3 AM, she had rewritten the pipeline. She imported the azureml.contrib module for the new data-drift detectors, slotted in the BanditPolicy , and linked the model to the Responsible AI dashboard.

Then she remembered the PDF.

She flipped to Chapter 12: Responsible AI . A case study mirrored her exact problem: biased sampling in a regional dataset. The author had included a code block for Azure’s ResponsibleAI dashboard, a tool she didn’t even know existed. It showed how to decompose a model’s error by subgroup.

She hit .