Contemporary autonomous systems increasingly operate through layered delegation: an orchestrating agent authorises a subordinate agent, which in turn authorises a tool. Each step may be individually sanctioned, yet the aggregate can produce authority structures that no designer intended and no audit log can reconstruct. This paper identifies chain integrity as the minimal structural primitive required to prevent that collapse — a boolean predicate over the delegation structure that either holds for an entire chain or fails the chain as a whole. We formalise four necessary conditions (scope containment, causal traceability, revocation propagation, and replay verifiability) and show they are jointly sufficient within the AMO governance model. We demonstrate that cascade revocation is a logical consequence of chain integrity, not an optional feature; that replay is a governance requirement, not optional tooling; and that chain integrity operates at a structural layer below policy — governing authority derivation rather than permitted behaviour. We include a comparative position relative to RBAC, ABAC, and capability systems; an independence argument for the four conditions; and an explicit treatment of threats to the theory. The AI-Delegation-Learning-Lab provides empirical validation.
Ricardo Rubio Albacete (Sun,) studied this question.