Myelith — A Decentralized Network in Which Consensus Work Powers an Agentic Language Model

Abstract

Proof-of-work blockchains purchase their security through the expenditure of compute and energy, yet the work performed is itself discarded entirely. Decentralized AI networks provide useful compute but secure no ledger. Myelith unifies both functions: miners jointly operate a large agentic language model via pipeline parallelism, and the same cryptographically attested inference work (“Proof of Inference”) determines compensation and feeds the voting weight of consensus. The native coin MYL closes the value cycle through a burn-and-mint equilibrium. The paper specifies a layered architecture that decouples consensus latency from inference latency; a verification model based on fully integer execution, where bit equality arises from the associativity of integer addition without any prescription over the order of operations, so that heterogeneous hardware participates without throughput loss; a token economy with a quantifiable security condition; a training procedure that verifies data provenance rather than data content and allows the network to grow its model incrementally; and an agent layer that makes the limit of verifiability at the boundary to the outside world visible. Changes from v0.1: the project was renamed from “Myelin” to “Myelith”. The verification model was rebuilt on integer arithmetic. New chapters cover training and model development, confidentiality with risk classes, and issuance structure; the agent layer was elaborated. Appendix B documents the rejected alternatives, supported by fourteen executable simulations. This design is not implemented and not measured; all figures rest on model calculations. No token exists. Eighteen open research questions are named as measurement tasks.