This paper addresses the long-standing problem of inconsistently defined decentralization in computer systems by introducing a formal, graph-based ontology. This framework defines decentralization as a structural, relational, and subject-specific property, distinguishing it from distribution and enabling multi-dimensional evaluation with new metrics like Void Tolerance and Imperviousness.

This paper proposes viewing large language model (LLM) teams through the lens of distributed systems to create a principled framework for their design and evaluation. It reveals that many established advantages and challenges from distributed computing, such as scalability limits and coordination issues, directly apply to and explain the behavior of LLM teams.