Time, Causality, and Observability Failures in Distributed AI Inference Systems

ArXi:2604.21361v1 Announce Type: new Distributed AI inference pipelines rely heavily on timestamp-based observability to understand system behavior. This work nstrates that even small clock skew between nodes can cause observability to become causally incorrect while the system itself remains functionally correct and performant. We present controlled experiments on a multi-node AI inference pipeline, where clock skew is