Geopol Forecaster

Open-sourced geopolitical prediction runs using a two-stage LLM ensemble pipeline

The Stack

The pipeline combines two established LLM ensemble techniques into a single grounded forecasting system. Each stage runs independently — when their probability estimates converge, it suggests genuine signal rather than averaging noise.

A

Stage A — Actor Simulation

Inspired by Snowglobe (IQTLabs)

Monte Carlo actor simulation with a referee + per-actor control loop. Each actor commits privately and independently. A referee narrates the evolving world state between turns using authority-precedence conflict resolution. Actors see only referee-authored state and their own private memory — no cross-contamination.

B

Stage B — Council Protocol

Inspired by LLM Council (Karpathy)

Six analytical lenses (Neutral, Pessimistic, Optimistic, Blindsides, Probabilistic, Historical) deliberate via a 3-stage protocol: parallel independent answers, blind peer review, chairman synthesis. No cross-contamination between lenses until review stage.

Infrastructure

LangGraph SQLite checkpointing Tavily search RSS / ISW feeds Pinecone vector memory OpenRouter Claude Sonnet 4.5 Typst PDF rendering
0
Gather & freeze live news
Tavily + ISW/RSS
A
Actor simulation
N actors × T timesteps
B
6-lens council
Answers → Review → Chairman
out
Report + PDF
Typst rendering

Experiment Design Philosophy

Both stages are grounded with a frozen news bundle (Tavily search + ISW/RSS feeds) so every actor and every council lens reasons from an identical world state. All LLM calls are routed through OpenRouter to the same model. Stage A produces emergent outcomes from N agents reasoning from their own persona in a game-theoretic loop. Stage B produces calibrated probabilities from expert-style analytical reasoning. Neither stage has access to the other's reasoning during execution. When they converge on similar probability estimates, it provides evidence the pipeline is surfacing genuine signal.

#3

Iran Forced Regime Change Probability

10 April 2026 · 10 actors · 6 lenses · Claude Sonnet 4.5

Question: What is the probability of forced regime change in Iran — defined as the IRGC being removed from power and replaced by a non-autocratic government that formally recognises Israel and renounces eliminationism — within 1 week, 1 month, and 1 year?

<1%+1 week
1-3%+1 month
4-9%+1 year

Forced regime change remains highly unlikely. Despite 85% defence industrial base destruction and 92% public dissatisfaction, the IRGC maintains institutional cohesion. The most probable pathway involves cascading failures requiring 9-18 months to materialise.

Per-Lens Probability at +1 Year

Simulation vs Council Across Horizons

Key Findings

  1. IRGC institutional resilience exceeds military degradation. Despite 85% industrial base destruction, IRGC controls 40%+ of economy and maintains internal security apparatus.
  2. Opposition fragmentation is the binding constraint. 92% dissatisfaction does not equal organised capacity. No parallel governance structures visible.
  3. Nuclear escalation continues. 60-75% probability of 90% uranium enrichment at Fordow within a month.
  4. Historical precedent unfavourable. IRGC survived Iran-Iraq War, 2009 Green Movement, 2019 protests, and 2022 Mahsa Amini uprising.
  5. Simulation overestimates regime fragility. Stage A estimated 12% at 1 year — discounted to 4-9% by the council based on fresh data constraints.

Experiment Parameters

Actors10 (Khamenei, Netanyahu, Trump, IRGC, Hezbollah, CENTCOM, Mossad, IDF, Russia, MBS)
Timesteps3 (+1wk, +1mo, +1yr)
MC runs1
Actor temp0.8 (exploratory)
Referee temp0.4 (consistent)
Council protocol3-stage (0.6 / 0.4 / 0.5)
Grounding6 Tavily queries + RSS (48h)
Chairman max tokens12,000
#2

Iran-Israel-US Ceasefire Durability

9 April 2026 · 38 actors · 6 lenses · Claude Sonnet 4.5

Question: What is the probability that the two-week ceasefire announced on 8 April 2026 between Iran, the United States, and Israel holds at +24h, +72h, +1 week, and +1 month?

55%+24 hours
22%+72 hours
10%+1 week
4%+1 month

The council assessed the ceasefire as a tactical pause, not genuine de-escalation, with rapidly decaying survival probabilities. The 72-hour mark (April 11-13) was identified as the critical inflection point.

Ceasefire Survival by Lens

72h Convergence — All Perspectives

Stage A / Stage B Convergence

The most notable experimental finding: Stage A (actor simulation) and Stage B (analytical council) converged to within 6 percentage points on the critical 72-hour window — despite being completely independent processes with no shared reasoning.

28%
Simulation (Stage A)
22%
Chairman (Stage B)
6pt
Spread

Key Findings

  1. Lebanon scope ambiguity is weaponised, not accidental. Both sides exploit undefined terms to continue operations while claiming compliance.
  2. Hezbollah is the most likely immediate trigger (24-72h window) — has never maintained operational pause >72h while under sustained attack.
  3. Hormuz status is the sole genuine de-escalatory action. Iran keeping the Strait partially open (10-15 vessels/day vs pre-war 135/day) as leverage.
  4. Russian evacuation from Iranian nuclear facilities was the single most important fresh data point, suggesting Israel planning strikes within 72-96h.
  5. Mojtaba Khamenei's functional status identified as highest-impact uncertainty — if incapacitated, ceasefire collapse rises to 80% within 72h.

Experiment Parameters

Actors38 (Iranian factions, Israeli security, US apparatus, Lebanon, proxies, regional, global)
Timesteps4 (now, +24h, +1wk, +1mo)
LLM calls~165
Wall clock~18 min
Estimated cost$6-12
Actor temp0.8
Referee temp0.4
Grounding4 Tavily queries + ISW/RSS (48h)

Accuracy Tracking

All predictions are graded on a standard 4-point rubric as prediction windows close. Wire services (Reuters, AP) take precedence over opinion sources.

Run 1 — Iran-Israel-US Conflict Trajectory (24/03/2026)

Run Score: 0.75 — 5 Correct, 1 Incorrect

#WindowPredictionOutcomeGrade
1.124h-1wk Israeli operations continue at full intensity IDF continued strikes through to April 8 ceasefire Correct
1.224h-1wk Iranian proxy retaliations continue Hezbollah maintained 55+ attacks/day Correct
1.31mo Iranian ballistic missile capabilities degraded 50%+ Coalition struck ~330/470 known launchers (70%) Correct
1.41mo IDF maintains/extends buffer zone to Litani IDF reached Litani River by March 24 Correct
1.51wk March 27 deadline fails No deal reached; war continued until April 8 Correct
1.61mo No ceasefire holds before March 27 Ceasefire reached April 8 — model assessed as improbable Incorrect

Runs 2 & 3 — Windows Still Open

Prediction windows for the ceasefire durability and regime change runs have not yet closed. Accuracy will be assessed as ground truth becomes available.

All Prediction Runs

#DateTopicActorsLensesModelKey Finding
310/04/2026 Iran forced regime change 106Claude Sonnet 4.5 <1% / 1-3% / 4-9% at +1wk/+1mo/+1yr
209/04/2026 Iran-Israel-US ceasefire durability 386Claude Sonnet 4.5 22% chance ceasefire holds at +72h (sim: 28%)
124/03/2026 Iran-Israel-US conflict trajectory --6-- 5/5 high-confidence predictions validated