The Prediction Process

This is an AI "remote viewing" experiment - can models trained on different data predict future geopolitical events with both accuracy and timing?

Daily Cycle

1. 23:00 UTC: Models read last 24h of defense news
2. Generate Predictions: Each model predicts 6-8 headlines for the next week
3. 7 Days Later: Compare predictions to actual headlines that appeared
4. Score: Dual embedding judges measure accuracy × timing precision

Scoring System

Each prediction is matched against actual headlines using:

• Semantic Similarity (40%) - Embedding-based headline matching
• Region Match (20%) - Geographic accuracy (Europe, Indo-Pacific, etc.)
• Actor Overlap (20%) - Country/organization accuracy
• Event Type (20%) - Category accuracy (cyberattack, sanctions, etc.)

Time Decay Penalty

Predictions earn less if they're late:

• Day 0 (predicted date): 100%
• Day 1: 90% • Day 2: 80% • Day 3: 70%
• Day 4: 60% • Day 5: 50% • Day 6: 40%

Consensus Bonus

When Western and Eastern teams independently predict the same event type + region:

• +20% bonus if the consensus prediction comes true
• Strong signal when models trained on completely different data agree
• Tracks "collective unconscious" of AI forecasting

Dual Judges

Predictions are scored independently by two embedding models:

• Western Judge: Nomic Embed (US/Nomic AI) - Western training data
• Eastern Judge: BGE-M3 (China/BAAI) - Chinese training data

Why? Embedding models trained on different corpora may interpret geopolitical text differently. Showing both scores + divergence reveals potential bias.

Predictor Models

• Western Team: LLaMA 3 (Meta), Gemma 2 (Google) - US/Western web training
• Eastern Team: Qwen (Alibaba), DeepSeek (CN) - Chinese web + government sources