Geopol Forecasts Index

April 18, 2026 · View on GitHub

Index of prediction runs across two related pipelines:

  • Geopol Forecaster — two-stage pipeline combining Snowglobe-style actor simulation with LLM Council-style multi-perspective analytical review.
  • Geopol Forecast Council — lean spin-off that skips actor simulation and polls a lineage-diverse LLM panel for concrete predictions per horizon over a grounded SITREP.

Each run is seeded with live news data and produces probabilistic or structured-prediction forecasts across multiple time horizons. Accuracy analysis is added as prediction windows close.


Prediction Runs — Geopol Forecaster

#DateTopicActorsLensesModelOpen-Sourced ReportKey Finding
310/04/2026Iran forced regime change probability106Claude Sonnet 4.5Iran-Regime-Change-Forecast-1004<1% / 1-3% / 4-9% at +1wk/+1mo/+1yr. IRGC institutional resilience exceeds military degradation; opposition fragmentation is binding constraint
209/04/2026Iran-Israel-US ceasefire durability386Claude Sonnet 4.5Iran-Israel-Ceasefire-Prediction-09042622% chance ceasefire holds at +72h (simulation: 28% -- 6-point convergence between independent stages)
124/03/2026Iran-Israel-US conflict trajectory (Day 25)--6--Iran-Israel-Conflict-Trajectory-2403265/5 high-confidence predictions validated; ceasefire (April 8) was not anticipated

Prediction Runs — Geopol Forecast Council

#DateTopicPanelHorizonsHead / AuthorRun DirectoryKey Finding
C118/04/2026Iran-Israel-US conflict evolutionGLM 5.1 · DeepSeek v3.2 · Gemini 3 Flash · Claude Sonnet 4.6 · Kimi k2.524h / 1w / 1mClaude Sonnet 4.6 / Claude Sonnet 4.62026-04-18_235440 (main PDF · full PDF)Panel confidence range 0.31–0.52 at 24h; escalatory predictions outweighed conciliatory at 24h. Modal views: nominal Lebanon ceasefire deteriorating, Hormuz closed, Hormuz/proxy-violence risks dominating short horizon

Accuracy Rubric

All predictions are graded on a standard scale. The /validate-predictions skill uses Tavily to search for real-world outcomes and applies this rubric automatically.

GradeScoreCriteria
Correct1.0Core prediction matched reality in direction and approximate magnitude/timing
Largely Correct0.75Direction right, magnitude or timing off by a modest margin
Partially Correct0.5General direction right but significantly off on timing, magnitude, or mechanism
Incorrect0.0Prediction contradicted by what actually happened
Not Yet TestableWindow still open or insufficient data to assess

Scoring rules: Direction is scored first, then adjusted for timing and magnitude. A >2x timing error drops Correct to Largely Correct or below. Binary predictions ("X won't happen") are scored on the stated claim, not the probability. Wire services (Reuters, AP) take precedence over opinion sources.


Accuracy Tracking

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

Run Score: 0.75 (5 Correct, 1 Incorrect out of 6 graded)

#WindowPredictionOutcomeGradeScore
1.124h-1wkIsraeli operations continue at full intensityIDF continued strikes through to April 8 ceasefireCorrect1.0
1.224h-1wkIranian proxy retaliations continueHezbollah maintained 55+ attacks/day through the periodCorrect1.0
1.31moIranian ballistic missile capabilities degraded 50%+Coalition struck ~330/470 known launchers (70%)Correct1.0
1.41moIDF maintains/extends Lebanon buffer zone to LitaniIDF reached Litani River by March 24Correct1.0
1.51wkMarch 27 deadline fails (no deal reached)No deal reached; war continued until April 8 ceasefireCorrect1.0
1.61moNo ceasefire holds before March 27Ceasefire reached April 8 — model assessed this as improbableIncorrect0.0

Run 2 — Iran-Israel-US Ceasefire Durability (09/04/2026)

Run Score: TBD (0 of 1 predictions graded)

#WindowPredictionOutcomeGradeScore
2.1+72h22% chance ceasefire holds at 72h (simulation: 28%)Not yet assessedNot Yet Testable

Pipeline Overview

Stage 0: Gather fresh data (Tavily news search + ISW/RSS feeds)
         Freeze into identical world state for all participants
              |
Stage A: Snowglobe-style actor simulation
         N actors x T timesteps, referee-narrated world state
         Each actor commits privately and independently
              |
Stage B: LLM Council 6-lens analytical panel
         Independent answers --> blind peer review --> chairman synthesis
         Sees: prompt + fresh data + Stage A simulation summary
              |
Output:  Chairman report, per-lens answers, cross-reviews,
         simulation transcript, executive PDF, archival PDF

What makes the convergence interesting

Stages A and B are completely independent analytical processes:

  • Stage A simulates actors making decisions in a game-theoretic loop -- it produces emergent outcomes from 38 agents reasoning from their own persona
  • Stage B runs six analytical lenses through structured deliberation -- it produces calibrated probabilities from expert-style reasoning

Neither stage has access to the other's reasoning during execution. When they converge on similar probability estimates (as in the 09/04/2026 run where the simulation gave 28% and the council gave 22% for the 72-hour window), it suggests the pipeline is surfacing genuine signal rather than averaging noise.



Author

Daniel Rosehill danielrosehill.com