Multi-actor LLM simulation for geopolitical forecasting, policy modelling, and contingency planning
Geopol Modeller is a fork of Geopol Forecaster by IQTLabs (In-Q-Tel). The original project demonstrated that every stage of a text-based wargame -- from scenario preparation through post-game analysis -- can be carried out by LLMs. Read the original paper here.
This fork rewrites the stack, adds a scenario/actor system designed for real-world crisis forecasting, and introduces prediction tracking with accuracy grading against ground truth.
Note: This is a different simulator from danielrosehill/Geopol-Forecaster, which combines the original Snow Globe wargaming methodology with an LLM Council model (multiple LLM "advisors" deliberating in structured rounds). This repo (Geopol Modeller) uses the pure multi-actor simulation approach — LLM agents with geopolitical personas interact through a LangGraph state machine, producing emergent outcomes rather than structured council deliberation.
Forked 12 April 2026 from IQTLabs/geopol.
| Component | Original Geopol Forecaster | This Fork (Geopol Modeller) |
|---|---|---|
| LLM access | LangChain + local models (llama-cpp, transformers/torch) | OpenRouter gateway (any provider via OpenAI SDK) |
| Model selection | Single model, configured in code | Model pools (YAML config, selectable at runtime) |
| Simulation loop | Imperative Python for-loop | LangGraph state graph (inspectable, pausable) |
| Pre-sim intelligence | None | Tavily web search + reference URL ingestion → SITREP |
| Actor system | Hardcoded in Python | YAML actor clusters with personas, red lines, constraints |
| Scenario system | Hardcoded | YAML scenarios with timeframe presets, assessment questions |
| Scenario subgraphs | None | Per-scenario graph configs (visibility, shocks, escalation) |
| Progress reporting | Print statements | Rich terminal UI with progress bars |
| Checkpointing | None | Auto-save after each move, resume from checkpoint |
| Report output | None | LLM-generated Typst PDF reports |
| Audio output | None | Edge-TTS podcast generation |
| Prediction tracking | None | SQLite DB with structured predictions, horizons, probabilities |
| Accuracy grading | None | Automated grading against real-world outcomes (Tavily) |
| Self-healing | None | Feedback loop: grade predictions → analyse variance → improve |
| Deployment | Docker (local) | Local CLI + Modal (serverless cloud) + FastAPI web dashboard |
| Dependencies | LangChain, torch, transformers, llama-cpp | OpenAI SDK, LangGraph, Tavily, Typst, Rich |
┌─────────────────────┐
│ Scenario YAML │
│ + Actor Cluster │
│ + Graph Config │
└─────────┬───────────┘
│
▼
┌───────────────────────────────┐
│ SITREP Agent │
│ Tavily search + ref URLs │
│ → military-format briefing │
└───────────────┬───────────────┘
│
▼
┌──────────────────────────────────────────────────────┐
│ LangGraph Simulation Loop │
│ │
│ ┌───────┐ ┌──────────┐ ┌─────────────┐ │
│ │ Setup │───▶│ Players │───▶│ Adjudicate │──┐ │
│ └───────┘ │ respond │ │ (narrator) │ │ │
│ │ (×N) │ └─────────────┘ │ │
│ └──────────┘ ▲ │ │
│ │ │ │
│ next move ───────┘ │ │
│ │ │
│ all moves │ │
│ done │ │
│ ▼ │
│ ┌────────┐ │
│ │ Assess │ │
│ └────────┘ │
└──────────────────────────────────────────────────────┘
│
┌───────────────┼───────────────┐
▼ ▼ ▼
┌──────────────┐ ┌────────────┐ ┌────────────┐
│ Prediction │ │ Typst PDF │ │ Podcast │
│ Extraction │ │ Report │ │ (TTS) │
│ + SQLite DB │ └────────────┘ └────────────┘
└──────┬───────┘
│
▼
┌──────────────┐
│ Accuracy │
│ Grading │
│ (vs. real │
│ world) │
└──────────────┘
The simulation avoids monolithic state actors. Internal dynamics drive real-world outcomes:
IRANIAN BLOC ISRAELI BLOC EXTERNAL
─────────── ──────────── ────────
┌──────────────┐ ┌──────────────┐
│ Khamenei │ │ Netanyahu │ ┌─────────┐
│ (Supreme │ │ (PM) │ │ Trump │
│ Leader) │ └──────┬───────┘ └────┬────┘
└──────┬───────┘ │ │
│ ┌──────┴───────┐ ┌────┴────┐
┌──────┴───────┐ │ Coalition │ │ CENTCOM │
│ IRGC │ │ Right │ └─────────┘
│ Command │ │ (Ben-Gvir, │
└──────┬───────┘ │ Smotrich) │ ┌─────────┐
│ └──────────────┘ │ China │
┌──────┴───────┐ └─────────┘
│ Basij │ ┌──────────────┐
│ (internal │ │ Opposition │ ┌─────────┐
│ security) │ │ (Gantz, │ │ Pakistan│
└──────────────┘ │ Lapid) │ │(mediator│
└──────────────┘ └─────────┘
IRANIAN OPPOSITION
────────────────── ┌──────────────┐ ┌─────────┐
┌──────────────┐ │ IDF General │ │ Turkey │
│ Street │ │ Staff │ └─────────┘
│ Movement │ └──────────────┘
└──────────────┘ ┌─────────┐
┌──────────────┐ ┌──────────────┐ │ Russia │
│ Silent │ │ Mossad │ └─────────┘
│ Majority │ └──────────────┘
└──────────────┘ ┌─────────┐
┌──────────────┐ │ MBS │
│ Pahlavi │ LEBANESE │ (Saudi) │
│ Diaspora │ ──────── └─────────┘
└──────────────┘ ┌──────────────┐
│ LAF │ ┌─────────┐
AXIS OF RESISTANCE └──────────────┘ │ UN SG │
────────────────── ┌──────────────┐ └─────────┘
┌──────────────┐ │ Gov of │
│ Hezbollah │ │ Lebanon │
└──────────────┘ └──────────────┘
┌──────────────┐
│ Houthis │ KURDISH
└──────────────┘ ──────
┌──────────────┐ ┌──────────────┐
│ Iraqi PMF │ │ Kurdish │
└──────────────┘ │ Factions │
└──────────────┘
Each scenario can define its own simulation behaviour without modifying
the core engine. Graph configs live in config/graphs/ and inject:
- Information asymmetry -- actors only see their own bloc's deliberations + public narrative, not adversary internals
- Exogenous shocks -- domain-specific unexpected events (e.g., "Hezbollah opens northern front") with weighted probabilities
- Escalation ladder -- structured 0-9 scale tracked alongside narrative output for quantitative cross-run comparison
- Adaptive tempo -- timesteps compress during kinetic escalation, expand during diplomatic phases
Scenarios without a graph config use the default generic loop unchanged.
# Install
pip install -e .
# Set API keys
export OPENROUTER_API_KEY=your_key_here
export TAVILY_API_KEY=your_key_here # optional, for current-events research
# Run with interactive model pool menu
geopol
# Run a specific scenario with a specific pool
geopol --scenario iran-israel-war --pool deepseek --report
# List available scenarios
geopol --list-scenariosPools define which models handle each role. Edit config/pools.yaml or pass a pool name at runtime.
| Pool | Planner/Narrator | Player/Advisor |
|---|---|---|
deepseek |
deepseek-v3.2 | deepseek-v3.2 |
anthropic |
claude-sonnet-4.6 | claude-haiku-4.5 |
google |
gemini-3-flash | gemini-3.1-flash-lite |
openai |
gpt-5-mini | gpt-5-nano |
xai |
grok-4 | grok-4.1-fast |
llama |
llama-4-maverick | llama-4-maverick |
minimax-mixed |
minimax-m2.7 | minimax-m2-her (roleplay) |
qwen |
qwen3.6-plus | qwen3.5-flash |
moonshot |
kimi-k2.5 | kimi-k2.5 |
zhipu |
glm-5 | glm-4.7-flash |
See docs/model-selection.md for benchmark links and guidance on choosing models.
Scenarios are YAML files in config/scenarios/. Each defines actors (inline or via reusable clusters in config/actors/), timeframes, assessment questions, and a nature parameter controlling narrator unpredictability.
Actor clusters follow a formal schema supporting personas, red lines, constraints, capabilities, internal factions, and influence models.
geopol --list-scenarios # see available scenarios
geopol --scenario iran-israel-war --pool deepseek --reportEvery simulation run automatically extracts structured predictions from the assessment phase and stores them in .geopol_data/predictions.db. Each prediction includes:
- Prediction text -- a specific, falsifiable claim
- Probability -- numeric confidence (0.0-1.0)
- Horizon -- time window (24h, 72h, 1w, 1m, 3m, 6m, 1y)
- Window opens/closes -- computed dates for when the prediction can be evaluated
Predictions are graded against real-world outcomes using a 4-point rubric aligned with the Geopol Forecasts Index:
| Grade | Score | Criteria |
|---|---|---|
correct |
1.0 | Core prediction matched reality in direction and approximate magnitude/timing |
largely_correct |
0.75 | Direction right, magnitude or timing off by modest margin |
partially_correct |
0.5 | General direction right but significantly off on timing/magnitude/mechanism |
incorrect |
0.0 | Prediction contradicted by what actually happened |
not_yet_testable |
-- | Window still open or insufficient data to assess |
geopol assess --all # grade all predictions with closed windows
geopol assess --run-id abc123 # grade a specific run
geopol predictions list # list stored predictions
geopol predictions summary # accuracy summary
geopol changelog # pipeline version historyThe self-healing loop is a feedback cycle that uses accuracy data to improve the simulation pipeline:
Run simulation -> Extract predictions -> Wait for windows to close
-> Grade against reality -> Analyse variance -> Suggest changes
-> Implement approved changes -> Run again
While the current focus is Iran-Israel conflict simulation, the engine is domain-agnostic. See planning/use-cases.md for the full roadmap, including:
- Geopolitical forecasting -- conflict simulation, crisis escalation modelling
- Foreign policy simulation -- sanctions testing, alliance stress-testing, treaty negotiation rehearsal
- Worst-case planning -- chokepoint disruption, nuclear escalation ladders, multi-front war pre-mortems
- International institutional modelling -- UNSC voting, OPEC+ dynamics, IAEA compliance disputes
- Domestic policy simulation -- noise regulation, housing policy, healthcare reform, transit planning
- Predictive forecasting -- election outcomes, ceasefire durability, treaty compliance monitoring
- Democratic resilience -- constitutional crisis red-teaming, power-sharing design
- Ideological lens analysis -- IR theory, economic schools, ethical frameworks applied as analytical agents
| Command | Purpose |
|---|---|
/run-sim |
Run a simulation on Modal cloud backend via MCP |
/run-scenario <name> [pool] |
Run a simulation locally via CLI |
/run-deepseek, /run-anthropic, /run-openai, /run-google, /run-xai, /run-llama |
Run with a specific model pool |
/run-all-pools |
Run the same scenario across all pools sequentially |
/post-run-analysis [run-id] |
Score a run's predictions against real-world outcomes |
/self-heal |
Analyse prediction variance and suggest pipeline improvements |
/backfill <scenario> [pools] [count] |
Seed the predictions DB with multiple runs |
/add-pool |
Add a new model pool |
/edit-scenario |
Edit the active scenario |
/check-models |
Check availability and pricing of models in current pools |
/research |
Research a topic and save findings |
This fork consolidates ideas and patterns from several prior experiments by the same author:
| Project | What It Contributed |
|---|---|
| AI-Agent-UN | 195-country agent roster, structured voting simulation, bilateral impact analysis |
| Peace-In-The-Middle-East | 150+ actor decomposition with sub-factions, chamber/working-group structure, facilitator agent pattern, faith and civil society actors |
| Panel-Of-Claude | Two-round deliberation (independent analysis then cross-pollination), tension mapping, motion decomposition, moderator synthesis |
| Claude-AI-Conference | Agent clustering by theme, identity generation for generic roles, broadcast mode (independent speeches with post-hoc synthesis) |
| Geopol-Forecasts-Index | Accuracy grading rubric, prediction tracking schema |
| AI-Ideation-Runs | Use-case brainstorming: 60+ ideas for multi-actor simulation across geopolitics, policy, economics, and humanitarian response |
- Original project: IQTLabs/geopol
- Paper: arxiv.org/abs/2404.11446
- Original authors: Daniel Hogan et al. (IQTLabs / In-Q-Tel)
Released under the Apache License Version 2.0.