Grounding Fabric Data Agents so they stop inventing columns | Part-2

Layer 5: Test for hallucination on purpose Everything above is guesswork until you measure it. <cite index="47-1">The Fabric SDK lets you define ground truth examples, run evaluations and analyse results in a notebook, so you can validate accuracy and debug errors before going to production.</cite> It is in preview. The standard approach is to write questions you know the answers to and check the agent gets them right. Do that. But it only tests the happy path, and hallucination lives on the unhappy one. Add negative test cases: questions whose correct answer is a refusal. These are the ones t