Recursive Semantic Divergence metric for LLM agent consistency

Recursive Semantic Divergence (RSD) provides an unsupervised trajectory-level metric to measure how agent behavior inconsistency propagates over time. Unlike token-level or single-response metrics, RSD uses bisimulation concepts from reinforcement learning to quantify cumulative semantic disagreement across multiple tool-use rollouts.