Documentation

Everything the benchmarks do, written down. Every protocol, every formula, every coverage figure, every version pin — public, and versioned by the evaluation run that used it.

That is not a nicety. A benchmark asks vendors to be measured in public, and asks clubs to act on the result. Neither is reasonable unless both can read exactly how a number was produced.

What is here

Start here explains what the Arena is and how to take part. Getting started for the shape of it, Quickstart to get an endpoint validated today.

Tasks has one section per benchmark. Transfers is the first, with its submission contract. Every task section has the same shape, so integrating a second one means reading a familiar layout rather than learning a new one.

Reference is the shared vocabulary — roles, metrics, and a glossary. Written once and used by every task, so a term cannot mean two things in two places.

Protocol covers the rules that apply across every benchmark — what is frozen, when predictions lock, what a refusal costs, what is published — and how to read a leaderboard, which is the page to send anyone about to act on a number.

Two rules that govern these pages

A version page is written once and never edited. Corrections are published as a new version with a changelog entry naming what they replaced. That is what lets a score from November still be readable against the rules that produced it.

Nothing here is generated by hand. Metric masks, band grids and sample payloads are rendered from the same specification a participant fetches from the API, so the page you are reading and the endpoint a vendor calls cannot say different things.

Status

This site is being built alongside the Transfers benchmark. Pages appear as their content is settled rather than as placeholders, so an empty section means the decision is still open, not that the page is missing.