Geopol Forecaster

A two-stage geopolitical forecasting pipeline. Given a forecast question and live news grounding, it simulates a roster of ~40 actors deciding how they would respond over four timesteps, then runs a six-lens analytical council that independently answers, cross-reviews, and synthesises a final report.

Every perspective is published — per-actor trajectories, per-turn world states, lens answers, cross-reviews — alongside executive-briefing and full-archival PDFs.

• Published runs: https://danielrosehill.github.io/Geopol-Forecaster/

Pipeline

1. Stage A — actor simulation. A referee + per-actor loop inspired by IQTLabs' snowglobe (examples/ac_sim.py). Every actor commits privately and independently; the referee narrates the world state between turns using authority-precedence conflict resolution. Actors see only the referee-authored state and their own private memory.
2. Stage B — lens council. Six lens directives (neutral, pessimistic, optimistic, blindsides, probabilistic, historical) deliberate via karpathy's llm-council 3-stage protocol — parallel query → blind peer review → chairman synthesis — with a frozen bundle of base context + Tavily/RSS/ISW fresh data + Stage A's simulation summary. The chairman writes the final report directly.

Grounding is a frozen bundle — no per-actor re-searching — so every member of the ensemble reasons from an identical world state.

Running

uv sync
cp .env.example .env     # fill in OPENROUTER_API_KEY and TAVILY_API_KEY
uv run geopol smoketest
uv run geopol forecast "Will the April 2026 ceasefire hold through +1 month?"

Each run lands in reports/<UTC timestamp>/ and is automatically rendered to three PDFs and fanned out to docs/runs/<id>/ for the Jekyll site:

• chairman_report.md — final synthesised forecast
• intel_report.pdf — executive briefing (~15 pages: cover + BLUF + chairman + colophon)
• full_transcript.pdf — archival (~300+ pages: every actor, turn, lens, review)
• combined_report.pdf — briefing + archival merged
• fresh_data.json — frozen Tavily + RSS/ISW bundle
• simulation.json — full Stage A transcript
• stage1_answers.md, stage2_reviews.md — council trace

Layout

geopol/
  config.py          Models, MC runs, paths.
  llm.py             OpenRouter async client.
  base_context.py    Static conflict background.
  schemas.py         Pydantic schemas.
  news/              RSS + Tavily + frozen fresh-data bundle.
  actors/            Persona briefs for the simulation roster.
  simulation/        Stage A — control/player loop + summariser.
  council/           Stage B — 6 lens directives + 3-stage protocol.
  render/            Markdown → Typst source → PDF.
  pipeline.py        End-to-end orchestrator.
  cli.py             `geopol forecast` / `geopol smoketest`.
scripts/
  render_intel_report.py      Executive-briefing PDF.
  render_full_transcript.py   Archival full-transcript PDF.
  publish_run.py              Fan out a run into docs/runs/<id>/.
  webui.py                    Stdlib web UI for submitting prompts.