swarm-rooms

GPU-accelerated multi-agent room simulation with CRDT merge and Eisenstein lattice compression.

Tests the hypothesis that swarm intelligence emerges from interconnection density, not individual capacity.

Key Concepts

• Rooms: N agents, each with small local context and large observation of others
• CRDT merge: observation updates commute (order-independent convergence)
• Deadband: only propagate when delta exceeds tolerance
• Eisenstein snap: compress observations to Z[ω] lattice points

Installation

pip install swarm-rooms[gpu]

Requires PyTorch >= 2.0.

Quick Start

from swarm_rooms import SwarmRoomNetwork

net = SwarmRoomNetwork(n_agents=256, device="cpu")
metrics = net.step()
print(metrics)  # propagation_rate, context_diversity, attention_entropy, ...

License

MIT