Geopol Modeller

Geopol-Modeller is a multi-actor simulation tool for geopolitical forecasting, policy modelling, and contingency planning. It is a fork of the IQTLabs Geopol Forecaster, which demonstrated that the stages of a text-based wargame can be carried out by large language models. The fork rewrites the stack around OpenRouter as a gateway to any LLM provider, uses LangGraph to drive the simulation as an inspectable and pausable state graph, and replaces hardcoded actors and scenarios with YAML-defined actor clusters, personas, red lines, and scenario configurations.

The pipeline begins with a SITREP agent that gathers pre-simulation intelligence through Tavily web search and reference URL ingestion, producing a military-format briefing. The simulation then cycles through player responses and narrator adjudication before an assessment stage. Outputs include LLM-generated Typst PDF reports, Edge-TTS podcast audio, and structured predictions stored in a SQLite database. Predictions are graded automatically against real-world outcomes, and a feedback loop analyses variance to refine later runs. The project is implemented in Python and includes a Rich terminal interface, checkpointing with resume support, and deployment options spanning a local CLI, Modal serverless execution, and a FastAPI web dashboard.

The project may be of interest to researchers in geopolitics and wargaming, analysts exploring scenario-based forecasting, and developers building LLM agent systems with stateful orchestration. Its documentation notes that it is distinct from a separate LLM Council variant, as it relies on emergent outcomes from persona-driven agent interaction rather than structured council deliberation.