Replication and Registry Accuracy in Blockchain RPC Evidence Policies: A Paired Synthetic Study

Blockchain applications often interpret agreement among RPC endpoints as evidence for a local decision. Endpoint multiplicity and source independence are distinct properties, and a policy that counts declared dependency domains relies on the quality of its registry. We present a reproducible paired synthetic evaluation of endpoint-counting and declared-domain-counting policies under shared context and expiry checks. An initial study covers four replication regimes and three thresholds using 2,000 shared seeds, producing 24,000 paired configuration cases and 48,000 decisions. A registry extension separates true source domains from policy-visible labels and evaluates splitting, merging, and unknown mappings at three perturbation probabilities. Across 1,000 shared seeds, the extension produces 120,000 paired configuration cases and 240,000 decisions. In the faulty-domain replication regime with threshold three, the domain policy produces 35 erroneous authorizations per 1,000 cases with an accurate registry and 166 when every true domain is split into endpoint-specific labels. Complete merging or removal of labels produces verification requests in all cases at that threshold. These findings characterize the interaction between replication, registry quality, and decision availability within the specified generator. The artifact provides executable controls, full trial outputs, and explicit assumptions for further RPC policy research.