The Testing Gap

Only 141 of 95,876 AI agents have a tested halt mechanism.

95,876
AI agents in the substrate
5,530
Plausibly tested halt
141
Documented tested halt

An untested halt is an assumption. Of 95,876 AI agents in 543 regulated financial institutions, 5,530 carry a halt mechanism described as plausibly or self-assertedly tested. Only 141 carry one with documented evidence of testing. That is 0.15 percent of the agent population. The remaining 99.85 percent operate on the presumption that a brake the institution has not exercised will work the moment it is needed.

Figure 1 · Funnel from Population to Documented Tested Halt
95,876 TOTAL AGENTS IN SUBSTRATE All AI agents identified across 543 institutions 100% of total HALT MECHANISM PLAUSIBLY TESTED halt_tested = likely or yes (self-asserted or inferred) 5,530 = 5.8% of total HALT MECHANISM DOCUMENTED AS TESTED halt_tested = yes (documented evidence of testing) 141 = 0.15% of total −94.2% lost: not even self-asserted as tested −94.5% lost: self-asserted but not documented Of every thousand agents, fewer than two have a documented test of their halt mechanism. MAR®500.com

The figure shows the three-stage funnel from the full agent population to the subset with documented evidence of halt-mechanism testing. Each bar is sized proportionally; the bottom bar is held at minimum visibility and remains barely discernible at full proportional scale because 141 sits against 95,876. An untested halt is an assumption made under regulatory pressure. DORA Article 11 and EU AI Act Article 14 require demonstrable, exercisable halt capability, not the existence of a procedure on paper.

Source · Meridian substrate v13.1.0, May 2026 · halt_tested classification · Documented = explicit yes value 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.