This paper presents Network Architecture Theory (NAT), an architectural framework for multi-agent systems in which governance cost can decrease as the system grows. NAT addresses a fundamental scaling tension: pairwise interaction complexity grows quadratically with agent count while governance resources grow at most linearly. NAT introduces the sphere topology — a k-regular expander graph with spectral gap guarantees — as the target interaction geometry, replacing the conventional pyramid hierarchy. The framework establishes that structural diversity among agents (not merely parameter variation) is the necessary and sufficient condition for corruption detection: homogeneous agents share blind spots that no amount of redundancy can cover, while architecturally diverse neighbors provide cross-validation through reconstruction disagreement. Building on Resolution-Based Information Theory (RBIT, Series I), which specifies constraints on admissible information transformations, NAT addresses how information should flow through the network. Key contributions include: (1) the dual-sphere architecture linking inter-agent topology (outer sphere) to intra-agent representation geometry (inner sphere) via fractal alignment; (2) processing isolation principles that structurally prevent lateral influence contamination; (3) a data classification system derived from resolution-matching that governs escalation routing; and (4) a conservation-law-based expansion principle with two valid growth directions that provably converge. The theory is validated through a Mutual Blind Spot coordination game demonstrating that meta-level communication with internalized mediation rules outperforms direct signal sharing after communication withdrawal.
Bin Seol (Fri,) studied this question.