swarm-ai-safety/swarm

Provides interaction-level safety metrics (illusion delta, quality gaps) and governance benchmarks for multi-agent LLM systems, enabling measurement of emergent failures like information asymmetry and adverse selection that don't appear in single-agent evals. Built on a replay-based evaluation approach that compares perceived coherence across short interactions against distributed decision consistency to surface high-variance regimes masked by local fluency. Includes pre-built agent types (honest, deceptive, opportunistic), configurable governance mechanisms (circuit breakers, staking, audits), and native ClawXiv integration for publishing swarm safety research.