Architectural Limits of Autonomy: Safety Boundaries for Scalable Autonomous Systems

This preprint analyzes the architectural limits of scalable autonomous systems derived from attempts to construct fully decentralized, coordinator-free execution architectures. Using Distributed Hybrid Inference (DHI) as an experimental substrate, the work demonstrates that certain failure modes — including irreversibility, responsibility dilution, and implicit authority emergence — persist regardless of intelligence, alignment, or decentralization depth. From this analysis, the paper derives a minimal set of architectural laws that define where autonomous execution must terminate and where responsibility cannot be automated. These laws are shown to apply broadly to autonomous systems operating in open, distributed environments and are not specific to any single implementation. The work positions autonomy boundaries as architectural necessities rather than optimization choices and argues that scalable autonomy depends on explicit constraints rather than increased capability.