The proof-of-concept could pave the way for a new class of AI debuggers, making language models more reliable for business-critical applications.