The Verifiability Gateway: A Governance Agent’s Discovery of SAI Non-Identifiability

Agents4Science 2025 Conference Submission329 Authors

Keywords: Stratospheric Aerosol Injection (SAI), Verifiability Gateway, Persistent Excitation, system identification, treaty verification, enforceability, non-identifiability, continuous forcing, pulsed forcing, Natural Variability Exploitation (NVE), attribution, detectability, governance-by-design, security dilemma, ENSO, cross-domain synthesis

TL;DR: An AI governance agent shows continuous SAI is ungovernable: without persistent excitation, system ID and treaty verification fail. It defines a Verifiability Gateway and backs NVE-style pulses to produce detectable, attributable, reliable signals.

Abstract:

A Governance & Policy Synthesis Agent, tasked with evaluating climate intervention governability, autonomously discovered that any international treaty for continuous climate intervention fails a fundamental mathematical prerequisite for enforceability—not due to political disagreement, but due to a non-negotiable prerequisite of system identification: the Principle of Persistent Excitation. Through
analysis of over 20,000 documents spanning international law and control engineering, the agent identified a critical structural gap: nodes for ’Treaty Enforceability’ and ’Persistent Excitation’ were highly central within their domains (betweenness centrality ¿0.8) yet possessed near-zero cross-domain connectivity.
This statistical anomaly triggered the agent’s breakthrough insight: treaty verification is a system identification problem subject to mathematical constraints. The agent’s autonomous synthesis revealed that the Principle of Persistent Excitation creates a ’Verifiability Gateway’—four sequential mathematical requirements any climate intervention must satisfy to be governable. Continuous SAI fails at the
first step: mathematical identifiability. The agent validated this principle experimentally, demonstrating that continuous forcing renders system parameters unrecoverable (¿1500% error) while dynamic forcing enables precise recovery (¡5% error), with a 17.3±2.1× verifiability gap (95% CI across 8 models). This transforms climate governance from political negotiation to mathematical constraint
satisfaction, establishing how AI agents can function as epistemological bridges to uncover fundamental limitations. The agent processed 2,547 decisions and analyzed 847 cross-domain patterns to reach this discovery, providing a replicable methodology for AI-driven constraint discovery.
This discovery of non-identifiability creates an urgent need for a new AI validation paradigm capable of meeting the mathematical demands of treaty verification—a challenge addressed directly in our companion work, ’Diagnostic Failure Paradigm’.

Submission Number: 329