Official Tests Accepted a Third of the Wrong Infrastructure Patches

Production AI infrastructure changes span distributed training, inference and reinforcement-learning systems, but most coding benchmarks stop at a kernel or one GPU. InfraDev reconstructs 36 pull requests from Megatron-LM, vLLM and verl as repository tasks that must run across many GPUs. Hidden probes cover requirements omitted by the original tests while an explicit prompt names every interface those probes invoke. Across six frontier models, the strongest resolved 38.9 percent. The probes rejected 32 percent of submissions that passed all official tests; using original pull-request descriptions as prompts produced no solved tasks. The benchmark is small and infrastructure-specific, yet it demonstrates why state, memory, time and traffic behavior need evaluation beyond nominal test suites.