Request Reuse and Consistency Boundaries in an EVM JSON-RPC Cache Proxy

An RPC cache can reduce repeated browser requests while changing the freshness and context of the answers a person sees. This report examines an EVM JSON-RPC proxy that combines method-dependent upstream selection, request coalescing, stale-while-revalidate reads, aligned log ranges, disk persistence and per-inner-call caching of Multicall3 aggregates. The evaluated module is an exact local working snapshot, distinguished by hash from the repository's committed version. Six original tests pass against loopback HTTP fixtures, and four newly authored diagnostics pass. In the original multicall fixture, the initial three client requests submit four distinct inner calls through two upstream aggregates; a later failing inner result is reused while a strict aggregate returns an error. Ten concurrent identical state requests produce one mock upstream request with distinct caller IDs preserved. Other diagnostics expose consistency limits: invalidation leaves an empty latest-code response cached, persisted results are reused after an upstream change, and the inner cache conflates different outer call contexts when mock responses depend on that context. These are local synthetic observations, not live RPC, EVM-contract execution, billing or latency benchmarks. The results support request reuse for the exercised read patterns while rejecting an unrestricted claim of transparent JSON-RPC equivalence. Empowerment requires visible freshness and context; affordability and useful financial access remain unmeasured; accountable operation requires preserving adverse results and the operator's cache policy. Jarett Dunn. Technical report, version 1.0. Preprint, not peer reviewed. Codex assisted inspection, diagnostic design, execution and writing. New report prose is CC BY 4.0. Original custom software retains its existing notices; no blanket code relicensing is asserted. Research helpers and upstream dependencies retain their stated licenses.