Beyond Delegation Depth: Audit Reach, Enforcement Reach, and Management Coverage in Recursive AI Delegation

Recursive AI systems increasingly permit agents to delegate tasks and authority to other agents. Contemporary control designs commonly use delegation depth as a prominent and easily enforceable guardrail: record the number of hops, impose a maximum depth, attenuate scope downstream, and propagate revocation through descendants. Those mechanisms are useful, but depth alone does not describe whether a recursive delegation structure remains manageable. This paper decomposes manageability into three distinct objects: path-level conditional management coverage C(p), summarized across depth by a coverage profile C_h; audit depth D_a; and enforcement depth D_e. Coverage records whether the management substrate survives across hops, while D_a and D_e record two specific functions of that substrate: reconstructability and authority invalidation. The quantities are defined relative to an explicit assurance context, so a guarantee is never interpreted outside its stated authority class, threat model, trusted components, and observation window. In branching delegation graphs, audit and enforcement reach are defined path by path; where reach is non-monotone, reach sets are reported rather than allowing a single deepest point to imply continuous coverage. In converging graphs, path-wise enforcement is necessary but graph-level revocation additionally requires closure across all independently sufficient paths, or across all minimal sufficient support sets under composite authorization, or an equivalent common enforcement point. The paper further establishes a sampling asymmetry as a proposition: audit sampling can support population-level evidentiary inference, whereas sampling alone cannot establish deterministic per-branch enforcement closure for an unsampled branch. The resulting policy conclusion is narrower than a ban on deep delegation: no single universal depth ceiling is sufficient as a general manageability criterion. Permissible depth should instead be conditioned on authority impact, reversibility, management coverage, audit reach, enforcement reach, alternate support paths, and revocation latency. The contribution is a conceptual measurement and governance framework; it does not claim novelty for recursive delegation, scope attenuation, maximum-depth controls, cascading revocation, audit sampling, or generic risk-graded oversight. Five falsifiable predictions (P1-P5) and a minimal factorial testbed are specified; no empirical results are reported in v1.0. This paper derives from the author's broader Information Ecosystem Theory, which remains unpublished at the time of this release; it is written to stand alone.