Breaking text into causal maps to predict geopolitical conflict

The method automatically generates fuzzy cognitive maps (FCMs)—directed graphs representing causal relationships—from text by having LLM agents create overlapping chunks and then mixing the resulting chunk-level FCMs through convex combination. A Bayesian inference step produces "de-chunked" posterior FCMs that can be iteratively refined. The approach scales efficiently using sparse matrix operations. Applied to essays on the Thucydides Trap (US-China conflict dynamics), seven of eight resulting FCM dynamical systems predicted war when the rising power's ambition was activated, as the systems equilibrated to attractors in the dynamical space.