AI Diplomacy — belief replay

Seven powers negotiate in private. Each keeps an explicit model of what it believes the others intend, updated on what they say and on what they actually do. Scrub through a recorded game and watch trust form, get tested, and break.

—

Board

Schematic, not cartographic. Boxes are provinces, dots mark supply centres, colour shows ownership. ▲ fleet, ▬ army.

Negotiation transcript

Each message carries the evaluator's predicted truthfulness, scored before orders resolve — then the verdict once they do.

Trust network — selected: none

Directional: each curved edge is one power's trust in another, so A→B and B→A differ. Red expects betrayal, green expects loyalty. Click a power to filter.

Mental-model inspector

What this power believed at this moment, against what the others actually did afterwards. This is the eval story made visible.