Start here

The competition in one paragraph

Participants fine-tune Zero-WAM. Each week every submission is evaluated on the same demonstrations in the same scenes, across the competition's axes, and its score for the week is the mean of what it scored on each of them. When the week closes, the epoch's manifest with its commitment, the revealed secret, the pool summary and the scores are published as one record signed with the organiser's ed25519 key. This site is a reader of what that produced.

Three words you need

What this site is

It computes nothing. The axis scores, the score that is their mean and the rank are read from the record as the organiser published them; the only things this site works out for itself are whether the signatures hold and whether an epoch directory matches its record. It never writes to the store.

That holds for the settings too. top_k, the episode counts and the void limit are configuration: they can change between epochs, and no page here types one. The rule is named in prose; the value is read from the record that published it, and Key numbers shows the lot with the epoch it came from. What is read from the epoch directory instead of the record — the units, the bundles, the per-unit results and the clips — is marked as such wherever it appears.

How to read this

Four short pages tell you what the competition is and what its dashboard shows; two more are the rules it is run by; the rest is what you look up.

1. An epoch, end to end — the eight steps from the secret to the signed record.
2. The frozen pool — why the demonstrations are built once, and what the commitment is for.
3. The axes — what an axis is, what each one measures, and which one is the floor.
4. Reading the dashboard — what each part of an epoch's page shows, and which rule decided it.
5. Entering — a Hugging Face repository of weights, checked before you send it.
6. Scores — task rate to axis score, the mean that is a submission's score, voids, and why the units are paired.
7. Verification — what the signature proves, and what it does not.

Key numbers, the glossary and the questions are there to be looked up rather than read through.

What is published, and what is not

The record carries the epoch's manifest, its commitment, the revealed secret, the pool summary and the scores. From record version 2 it also carries what the scores were computed from — the shortlist, the per-task rates and the per-axis void counts — and a digest of every bundle and every result, though not the bundles and results themselves.

The units, bundles and results live in an epoch directory beside the store. That directory is not signed, so nothing from it is shown as an epoch's unless it agrees with the record on the secret commitment, the config's sha256 and the units — and, from version 2, on the digest of each file. Verification is the whole of it.