Abstract
Recent formal work on human-agent collectives identifies an Accountability Horizon, a computable autonomy threshold above which single-locus accountability cannot satisfy a set of minimal axioms simultaneously (Tibebu, 2026). This impossibility result addresses collectives, meaning systems in which multiple agents act jointly within cyclic interaction structures. We address a structurally distinct and practically dominant architecture: the sequential delegation chain, in which a human principal delegates a task to an AI agent, which sub-delegates to another, producing a directed acyclic sequence of delegations before any action with real-world consequence is taken. We show that the Accountability Horizon result does not apply to acyclic delegation chains, and that in this regime a formal specification for chain accountability is both possible and achievable. We develop this specification by importing three centuries of common law agency doctrine into the delegation chain setting, producing five structural invariants for what we call a mandate-preserving accountability chain: Mandate Boundedness, Authority Narrowing, Scope Monotonicity, Forensic Reconstructibility, and Firebreak Completeness. We establish two main results: the Mandate Monotonicity Proposition, showing that scope and delegation authority are non-increasing along any well-formed chain; and the Accountability Firebreak Theorem, identifying conditions under which accountability is recoverable at a designated node following a scope violation. We show that these conditions correspond to concrete design requirements of verifiable delegation protocols. As a constructive existence witness, we demonstrate that the Tensflare Accountability Protocol (TAP), an open verifiable mandate architecture, satisfies all five invariants.