In a dbt interview, Shreya Shankar explained why real analytical work spans databases, unstructured information and domain knowledge. DataAgentBench covers 104 questions across 17 datasets, four database systems and six industries. The interview reports a best pass@1 of 61%, while the reviewed paper abstract reports a best success rate of 57%; those contexts should not be treated as one identical measurement. A semantic layer can help, but checking final SQL alone misses task-interpretation failures. Rubrics need inspection of real examples and supporting evidence.