Modeling systems before you have to live in them.

Some systems you never get to run an experiment on. A region under sanction, a supply chain mid-disruption, a market pricing a policy nobody has announced yet — the cost of finding out the hard way is the whole reason you are asking. The method is the same in each case: build the system out of agents that hold real positions, apply the pressure you are worried about, and watch what the agents do to each other. Two of ours are geopolitical, and they do different jobs.

AI World is the world model. Roughly two hundred countries are each played by a language model given that country's actual position — its economy, its alliances and dependencies, its government and leadership, and what it has already lived through. You define the scenario; each round, every agent decides what its country does in response, drawing on its persona, its memory of this run, and its standing doctrine. Nothing is scripted, so the same scenario run twice does not have to end the same way — which is exactly why you run it more than once.

Geo Forecaster is what running it more than once becomes. It freezes the evidence behind a hash so a re-run is grounded identically, proposes a small set of mutually exclusive outcomes, runs the same cascade engine repeatedly across varied conditions, and classifies where each run lands. A committee of specialist agents — base rates, status quo, devil's advocate, regional expert — then reads every run before probabilities are aggregated. What comes back is a probability per outcome with confidence intervals and its uncertainty decomposed, and the harness is scored against historical events whose outcomes are already known.

Both of these are geopolitical because that is where we built first, not because the method is. The same construction — agents holding real positions, a shock you specify, and the cascade that follows — applies to a supply chain, a market reacting to a rule that has not landed yet, a regulatory regime, or an organisation responding to its own restructure. If the system you need to rehearse is not on this page, that is a build, not a different discipline.

The first-order effects are rarely the interesting part; most rooms can already name them. The exercise is for the third and fourth step, where one actor's entirely reasonable response is a shock to somebody who was never on the map you drew. What you keep is a rehearsed set of ways a decision can go, an explicit map of who turns out to be exposed, and — when the question needs a number rather than a narrative — a calibrated probability with its uncertainty stated, arrived at before the decision is one you have already taken.