Iran Conflict Scenario Monitor

About This Exercise

Better than headlines, less than intelligence. This is a methodological demonstration, not an intelligence product. We are not Iran experts. The St. Gallen Endowment's expertise is in trade and digital policy, where we operate the Global Trade Alert and the Digital Policy Alert — two databases tracking over 85,000 government policy interventions worldwide. Every entry in those databases is verified by a trained human analyst. That human-verified evidence is our brand and the foundation of our credibility. We have no intention of replacing it with machines.

What we are exploring is how structured AI systems can extend the reach of human expertise, not substitute for it. In our core work, we are developing a two-tier model: human-verified evidence remains the authoritative layer; machine-monitored intelligence, clearly labelled as such, extends coverage into domains and geographies where full human curation is not feasible. The combination of both is what we believe produces the best coverage.

This monitor is an exercise in the second tier. It uses 14 specialised agents, each grounded in a specific theoretical framework — Schelling on bargaining structure, Posen on inadvertent escalation, Tetlock on forecasting discipline, and others. The agents assess evidence independently, challenge each other through a dedicated red team, and produce structured disagreements that force analysis beyond news summarisation. No human analyst has overridden the machine findings. The system's discipline comes from its architecture, not from human review of every output.

The analytical value is not in any single prediction. It is in the governance layer: how the agents are specified, how they interact, what decision rules constrain them, and how adversarial challenge is enforced. That governance structure is where domain expertise lives. The person who builds the agents encodes their knowledge into the system's rules, frameworks, and evaluation criteria. We are building the best agents in the world for trade and industrial policy monitoring, because that is where our fifteen years of classification experience can be encoded. This exercise shows what the architecture can do in a domain where we have no special advantage.

Framework attributions (Schelling, Posen, etc.) identify the analytical lens being applied, not empirical claims. Where evidence is thin, we say so.

This work builds on capabilities developed through our partnership with the Patrick J. McGovern Foundation and Stanford's Institute for Human-Centered Artificial Intelligence, where we are building an AI legislation tracker using the same methods — there, with our deep domain expertise providing the human layer that this exercise deliberately lacks.