Question
An AI agent rewrote a slow analytics query and a backing SQL migration, claiming the new query "returns identical results, just faster" by replacing a correlated subquery with a window function and adding an index. The numbers in a spot-check on staging matched. How do you verify 'identical results' rigorously before this replaces a query that feeds finance dashboards, and what's the failure mode you're specifically guarding against?
Treat the AI’s output as a draft to verify, not an answer to trust. Name the specific flaw and the input that triggers it, say how you’d catch it — tests, edge cases, reading critically — and how you’d re-prompt or decompose to get it right.
Vibe coding: describe the solution in plain language (or narrate it) and the coach grades your approach. Generating runnable code from your description is coming next.