Rules

Competition period

Submissions are accepted from 5 October 2026, 00:00 AoE to 31 December 2026, 23:59 AoE. AoE is Anywhere on Earth, UTC−12. A submission arriving outside that period is refused.

Prizes

The preview competition awards no prizes.

Teams

A team competes under a single participant reference, used on every track it enters. A person belongs to one team at most, and a team registers one reference at most. Teams have no size limit, and an individual competing alone counts as a team.

Submitting

Each track takes two files. The first is a CSV matching the format declared on that track’s page: exact columns, exact row count, unique identifiers. It is validated before it reaches the competition, so a malformed file never consumes a grading slot. A file can be checked without being submitted.

The second is a ZIP of the code that produced those predictions, at most 50 MB. The check confirms that it is a ZIP and within the limit, and does not read its contents. The archive is never published: it appears on no leaderboard and no submission page. The organizing team reads it to verify a submission and to understand the strategy adopted.

Automated Essay Scoring: 24 rows, identifier essay_id.
Mathematical Exam Diagnostic: 101 rows, identifier equation_id.
Answer Sheet Detection: 26 rows, identifier sheet_id.

Each team may make up to 10 submissions on each track, all under its one participant reference. A submission refused before it is accepted, by the check or by the competition, does not count toward that limit. An accepted submission counts whether it is scored or fails grading.

External data and pretrained models

Publicly available data and pretrained models may be used, including models reached through a commercial API. The code archive names each one, with a link and a version precise enough to identify it. Data that is not publicly available may not be used.

Prohibited conduct

Labelling the test items by hand, or having anyone else label them, whether the labels are submitted or used for training.
Recovering the test targets by any other means, including by inference from the scores of repeated submissions.
Using more than one participant reference for the same team, including to exceed the submission limit.
Sharing predictions or code with another team during the competition.
Submitting predictions that the code in the accompanying archive did not produce.
Interfering with this site, the competition service, or another participant’s submissions.

The organizing team may remove from the leaderboard any submission or participant that breaks these rules.

Ranking

A board carries one row per participant: that participant’s best scored submission, judged by the track’s ranking metric. A submission still being graded, or one that failed, takes no part and never displaces a scored one. Tied participants share a rank and the next rank skips the numbers the tie consumed, so a board reads 1, 2, 2, 4.

Baselines produced by the organizing team are submitted like any other entry, are ranked alongside everyone else, and are labelled on the board.

Participant identity

A participant reference is not authenticated.

A participant is created the first time a reference is used, and nothing checks who uses it afterward. The contact email is compared against what was recorded for that reference, across all three tracks, and a mismatch is refused. The check catches typos and accidental collisions. It does not authenticate anyone.

A reference whose submissions have all failed grading appears on no board, so nothing is recorded to compare against, and a later claim on it passes. A random string of letters and digits is harder to claim than a team name or team1. Leaderboard standings are provisional.

Display name and affiliation are fixed on a participant’s first submission, on any track, and cannot be changed from here afterwards. The email address is never published and never leaves the server.

Keeping a submission reference

Every accepted submission gets a reference, which is the only way to look that submission up. My submissions lists the references recorded in this browser. The list is kept in the browser’s own storage and appears on no other browser or device. Losing the list costs nothing as long as the references themselves are kept.

Changes and early closure

The organizing team may amend these rules, suspend submissions on any track, or close the competition before the scheduled date, at any point and without prior notice. Amendments are published on this page. A track that stops accepting submissions says so on its page, and its leaderboard remains readable.