A prototype shows how to investigate bugs more intelligently when AI-generated code passes tests but still fails in production
