The Scope Gap
Related
Cascade8.9% sub-min Cascade0 / 56,308 CascadeGCI 10.15

Most halt mechanisms stop only the agent that originated the failure.

36.8%
Agents with agent-only halt scope
8.7%
Agents with institution-wide scope
44.8%
Agents with no classified scope

Of 95,876 AI agents in 543 regulated financial institutions, 36.8 percent have a halt mechanism whose scope reaches no further than the agent itself. 9.7 percent can halt a wider business line. Just 8.7 percent have a halt that reaches institution-wide. The remaining 44.8 percent have no classified scope at all. A halt that stops one trading bot but not the cascade behind it is not containment.

Figure 1 · Halt Scope Distribution Across 95,876 Agents
UNKNOWN 42,935 · 44.8% no halt scope classified AGENT-ONLY 35,260 · 36.8% halt stops only the agent itself BUSINESS LINE 9,310 · 9.7% halt stops a cluster of agents INSTITUTION-WIDE 8,371 · 8.7% halt cascades across the institution stops the cascade ↑ stops nothing ↓ More than a third of agents can only stop themselves. MAR®500.com

Each band shows the share of agents with a halt mechanism at that scope. The widest band (UNKNOWN) holds the agents with no classified scope. The next (AGENT-ONLY) holds those whose halt stops only the originating agent. BUSINESS LINE halts can stop a cluster; INSTITUTION-WIDE halts can stop the cascade. Cascade containment requires reach beyond the originating agent. On the substrate as published, more than a third of agents have halts that stop only themselves, and only one in twelve has a halt that can stop the cascade.

Source · Meridian substrate v13.1.0, May 2026 · halt_scope field · All evidence tiers Methodology · Cascade Without Containment (working paper)
543 institutions · 12 sectors · 95,876 agents · 636,854 governance edges · substrate v13.1.0
Methodology grounded in Cascade Without Containment, working paper. Substrate methodology in The Stationary Sea (Part 1) on Zenodo.