Radix Research

Introduction

Most of Radix exists because one problem was refused for thirteen years. A single-threaded ledger cannot serve the world’s financial traffic, and the usual answers – bigger blocks, faster blocks, execution moved off the ledger – buy time rather than remove the ceiling. The research collected here is the record of the designs that were tried against that ceiling, including the ones that were abandoned.

Reading it in order is the fastest way to understand why Radix looks the way it does. The discarded work is not a footnote to the shipped protocol; it is the reason the shipped protocol has the shape it has.

The line of designs

eMunie is where it starts, around 2013 and before Ethereum launched, as Dan Hughes’ first distributed ledger project. Tempo, proposed in 2017 as the fifth iteration, was a DAG-based design that pioneered ledger pre-sharding and lazy consensus – the idea that related transactions can be grouped and that not every node needs to agree about every one.

Cerberus replaced Tempo and is the design Radix is named for in the literature. Its whitepaper appeared as a preprint in 2020 and was peer-reviewed and published in the Journal of Systems Research in June 2023, which is a rarer credential in this field than it sounds. Consensus evolution at Radix traces the whole sequence in one place.

Cassandra is the sharp-edged one. Hughes ran it to attack two questions a sharded network cannot avoid: what happens when a subset of validators stops processing transactions and liveness fails, and how validator sets can be reconfigured while the network runs.

Hyperscale, and what the tests measured

The current line is Hyperscale. In January 2026 the Radix Foundation completed a public test in which its reference implementation sustained more than 500,000 transactions per second and peaked above 700,000, on commodity AWS m6i.xlarge instances of four cores and 16 GB each. The test page carries the setup in full.

What the figure measures is worth stating precisely, because it is easy to over-read. The test exercised the Foundation’s Hyperscale implementation. It was not a measurement of Cerberus as specified, and the braided cross-shard consensus described in the Cerberus paper has not shipped in any implementation.

hyperscale-rs is the community-built Rust implementation of the Hyperscale protocol, led by flightofthefox of proven.network and opened for public review, and it is the leading candidate to deliver Xi’an. The economic model sits alongside this work, governing where XRD comes from and where it goes.

Why the record is kept

A protocol that has rewritten its consensus layer five times invites an obvious question about whether the sixth will land. Keeping the discarded designs legible is the honest answer to it: each iteration names a failure mode that the next one had to survive, and the current work can be judged against that list rather than against a promise. Xi’an is where the thirteen years get their result.