What happened
Researchers built a tool that tests whether AI-written SQL queries actually work correctly by finding cases where they fail. Instead of just checking outputs against examples, the tool mines real database constraints from actual data and uses an AI language model to validate them, making tests reflect what actually matters in practice.
Why it matters
Most AI systems that translate English to database queries are tested on standard benchmarks that miss real-world failures — this tool exposes those missed errors, which matters because faulty database queries in production silently return wrong answers instead of failing loudly.