Topology-Level Containment with Autonomous Withdrawal

We investigate whether recovery in adaptive multi-agent systems can emerge from topology-level intervention alone, without modifying agent states. A compact three-component adaptive buffering policy—coupling attenuation, environment-designed bridge rewiring, and temporary fatigue shielding—is tested across 80-node networks under three topologies (small-world, scale-free, Erdős–Rényi). Systematic ablation reveals clean functional separation between a containment engine (94.8% of DSI reduction) and a withdrawal engine (sole enabler of autonomous exit events). The policy achieves topology-invariant disorientation containment (ΔDSI ≈ −0.040, 9/9 wins) with cyclic engagement–withdrawal dynamics. Results are interpreted through the Deficit-Fractal Governance (DFG) framework as an internal mechanism decomposition study. Structural parallels to homeostatic synaptic scaling and bistable seizure dynamics suggest the identified regulatory motifs may belong to a broader class of stability–plasticity regulation in adaptive systems.