What It Is
Polyphony is a multi-model geopolitical simulation. Instead of running every actor on the same LLM, it assigns different models to different state actors based on their reasoning style — Iran on DeepSeek, the US government on Grok, Israel on Claude Opus, and so on.
The core question: do different models exhibit distinct “textures” in how they reason, and does that diversity produce more realistic simulations than a single model talking to itself?
The 建前/本音 Architecture
Every state actor runs a dual stack — a public-facing agent (建前) that issues press releases and UN statements, and a private deliberation agent (本音) that runs war cabinet reasoning visible only to back-channel partners.
The gap between what actors say publicly and what they’re actually planning is where the interesting signals live.
What Came Out of It
The first full run (Strait of Hormuz crisis scenario, March 2026) produced some genuinely surprising results:
- Iran’s public “absolute sovereignty” posture vs. private “need an exit within 4-6 weeks” — which matched real diplomatic dynamics
- The US government agent invented justifications and branded developments as personal victories without being prompted to
- The simulation correctly inferred several real-world developments 24-48 hours before they happened
Still validating the hypothesis against prediction market prices, but the early results are encouraging.