Replicating Belief, Not Bits: Epistemic State Replication for Agentic Systems

This paper argues that classical State Machine Replication is a poor fit for distributed systems built from stochastic, model-driven agents. It proposes Epistemic State Replication (ESR), which separates an immutable evidence log (L) from a stochastic belief lineage (B). ESR introduces Semantic Linearizability, Bounded Eventual Coherence, structured epistemic deltas, and Verifiable Semantic Rollbacks. The authors report preliminary simulation results suggesting feasibility under their stated assumptions and illustrating fewer secondary cognitive faults. The abstract does not provide detailed benchmark configurations, datasets, numerical results, or comparisons with specific baselines.