Official Rules

1. Agreement

By registering for or submitting to the Virtual Embryo Challenge (the “Challenge”), you agree to these Official Rules and to the Website Terms of Use, which form part of them. If you do not agree, do not register or submit.

The Challenge is organised by the Qiu Lab at Stanford University together with its collaborators (the “Organisers”). The Organisers may amend these Rules; material changes are announced on this page and carry a new version, and continued participation after a change constitutes acceptance of it.

2. Who Is Eligible To Enter

Each team must designate one captain who is at least 18 years old, or the age of majority in the captain’s jurisdiction if that is higher, at the time of registration. Other team members are not subject to a minimum age requirement under these Rules, provided their participation is permitted by applicable law and by any school, employer, institution, parent or guardian whose permission is required.

Only people who have made a genuine contribution to the team’s work may be listed as team members. Honorary, courtesy, placeholder and purely administrative membership is not permitted.

The following may not enter, and entrants may not use their assistance:

• Members of the organising committee, and anyone with access to the held-out evaluation data.
• Employees, interns and students of the organising labs — the Qiu Lab at Stanford, and Professor Neil Chi’s group at UCSD, which generated the datasets.
• People advised formally or informally by members of those labs during or after January 2026.
• Service providers, consultants, vendors and contractors to the Organisers, and their employees.
• Anyone whom applicable law prevents from participating, or to whom the Organisers are legally unable to award a prize. This is a constraint the Organisers are subject to, not a judgement about any country or institution.

Entry is void where prohibited. You are responsible for confirming that your employer or institution permits your participation.

3. Participating As An Employee Or University Representative

At registration you state whether you participate in a personal capacity, as an employee of an organisation, or as a representative of a university.

• A team whose members participate as employees must draw them from a single organisation. A team of university representatives may span more than one university.
• A team may not mix capacities — some members as employees or representatives and others in a personal capacity.
• Registering in an employee or representative capacity confirms you are authorised to enter on that organisation’s behalf, and “you” in these Rules then also means that organisation.
• Tell us immediately if that authorisation is withdrawn. The Organisers may verify authority directly with the employer or university, and may disqualify you or your team if it is absent.

4. Your Responsibility To Others

• Your participation must not breach any obligation you owe anyone else, including an employer, a client, or a school.
• You may not use property, know-how, code, information or data that you do not have the right to use for this purpose.
• You may not tamper with the administration of the Challenge, submit entries whose purpose is unrelated to it, collude with other teams in a way that undermines its integrity, or attempt to access another team’s work.

5. Fees

There is no entry fee. You are responsible for your own costs, including compute.

6. Registering

Registration requires a verified sign-in, your legal name, a contact email, and an affiliation. Affiliation may be “Independent researcher”; it may not be left blank or filled with placeholder text.

One account per person. One person may belong to only one team. Registering on behalf of an organisation warrants that you are authorised to bind it to these Rules.

Each member of a team accepts these Rules individually — a captain cannot accept on a teammate’s behalf, and each acceptance is recorded against the version in force at the time.

7. Teams

• A team may have up to 10 members, including the captain.
• The captain may invite members, remove members, and transfer the captaincy. Removing a member does not remove the submissions they made; those remain the team’s.
• Members may join a team at any time before the final submission deadline.
• Prizes are awarded to the team, and it is the team’s responsibility to divide them.
• Teams form and manage themselves: choosing members, assigning roles, documenting contributions and dividing prizes are all the team’s own. The Organisers do not form teams and are not responsible for disputes about team formation, membership, contribution, authorship or prize-sharing, except so far as administering or enforcing these Rules requires it.

8. Tracks

A team enters one of two tracks and competes only within it. The prize pools are separate and equal.

• Human Team — a conventional workflow: people design, train and select the method.
• Agent Team — the method is developed by an autonomous agent. A person may write the starting prompt, but nobody may read intermediate results and steer the agent on the strength of them.
• A captain may change track before the final deadline. Doing so retires every score the team has earned: the entries keep their numbers in the team’s own history but stop counting on every board, and the team starts the new track from nothing. Entries are judged under the rules of the track they were submitted in, and a result cannot follow a team across.

9. Agent Team Entries

This section applies only to teams in the Agent Team track. It states what makes an entry an Agent Team entry, and what has to be filed alongside it.

The distinction the track rests on is autonomy, not tooling. A person may write the harness, choose the model, write the initial prompt, set the budget, and restart a run that crashed. What a person may not do is read the agent’s intermediate results — scores, diagnostics, partial outputs — and steer the next step on the strength of them. An entry produced that way is a Human Team entry, whichever track it was submitted in.

• Every Agent Team submission must be accompanied by evidence of the run that produced it. A submission is accepted and held, but is not scored and does not appear on any board, until at least two distinct kinds of evidence have been uploaded against it. It scores automatically once they land.
• The kinds are: the trajectory (what the agent tried, what it observed, what it did next), the prompts (every prompt given to the agent, including the initial one), the harness (the code that ran the agent — orchestration, tools, evaluation loop), and other material needed to follow the run end to end. Any two distinct kinds satisfy the minimum; more is better and none is a substitute for another.
• Evidence is capped at 200 MB per file and 600 MB per team in total. Trajectories compress well; a trace that does not fit is a trace that should be filed in full and summarised, not truncated silently.
• Evidence must correspond to the run that produced the submitted file. Evidence from a different run, edited to fit, or reconstructed after the fact is a misrepresentation and is treated under Conduct and Disqualification.
• Predictions must be written by the agent. A file the agent produced and a person then edited — by hand, by script, or by substituting a better result from elsewhere — is not an Agent Team entry.
• A final entry nominated for prizes must carry its evidence. An Agent Team entry without it may hold a leaderboard position but is not eligible for a prize, and the Organisers are not obliged to ask for the missing files before the deadline.
• The Organisers may audit any Agent Team entry: read the evidence, ask questions about the run, and ask for the harness to be run again on request. Refusing an audit, or an audit that cannot reconcile the evidence with the submission, is grounds for disqualification.
• Evidence is held privately and used to verify eligibility and to report aggregate findings about how agents performed. It is not published, and it is not shared with other participants. Ownership, Licence and Publication governs anything the Organisers do publish.

If you are unsure whether a particular kind of human involvement crosses the line, ask before submitting rather than after. The Organisers would rather answer the question than adjudicate it.

10. Data and What You May Use

Competition data is released under the Terms of Use, clause 14, and is provided for research and educational purposes.

• Training stages are released at registration. Validation and test targets are never distributed while they rank: a submission comes back as a score, not as answers.
• External public data, pre-trained models and published code may be used, provided their licences permit it and their use is described in the method summary. For Task 1 this is expressly encouraged.
• Every external source must be disclosed with the submission. An undisclosed source is a rules violation whether or not it changed the result, because the disclosure is what makes the previous condition checkable.
• No measured data from a held-out stage or genotype may be used, by any route — not directly, not through a model pre-trained on it, and not through a public dataset that contains it. This is the one restriction that decides whether a result means anything.
• The held-out stages and genotypes are: E10.5 and E12.5 for Task 1; E7.5 and E7.75 in the embryo setting and E8.5, E10.5 and E12.5 in the heart setting for Task 2; and the Gata4 and β-catenin knockouts at E8.75 for Task 3. Everything else is fair game, and for Task 1 external data is expressly encouraged.
• Nearness counts, and the stage label is not the test. External data is treated as a held-out stage when it sits closer to that stage than to the nearest stage we released for the same task — the midpoint between them is the boundary. Holding out E9.5 and training on an external E9.45 embryo is using the answer under a different name; so is a dataset labelled with a somite count or a Theiler stage that lands in the same window. Where a held-out stage lies beyond the last released one, as the extrapolation targets do, everything from the midpoint onwards is treated as held out.
• For the extrapolation targets the window is stated outright rather than derived: no external data from any time point between E9.5 and E13.5 may be used, and E10.5 and E12.5 — the validation and test targets themselves — are excluded absolutely. Time points outside E9.5–E13.5 may be used, provided the source and its stages are stated explicitly with the submission.
• For interpolation targets, external data too close to a held-out stage is also treated as held out. In Task 2, this means heart data from E8.25 to E8.75 and embryo data from E7.5 to E7.75 may not be used. These intervals are the midpoint-derived protected windows around the held-out interpolation stages.
• The same applies to genotype. Data from the held-out knockouts, from another allele of the same gene at a comparable stage, or from a perturbation that phenocopies them, is treated as the held-out condition.
• A general-purpose resource that happens to span a held-out stage is not banned outright — the cells inside the window must be removed before training, and the removal described in the method summary. Using the resource with them left in is using the held-out data.
• If a source sits near the boundary and you are unsure, ask before submitting. The Organisers will answer, and an answer given in advance is not a finding of misconduct afterwards.
• Attempting to recover held-out labels, probe the evaluation infrastructure, or otherwise circumvent fair evaluation is prohibited and is grounds for disqualification.

11. Leaderboard Entries and Final Entries

Each task is scored on its own boards, and a submission names the board it answers. Limits are per team, per task, per UTC day:

• P1 · Development — 20 scored submissions per day.
• P2 · Baselines released — 8 scored submissions per day.
• P3 · Final test phase — 2 official submissions per board, for the whole phase rather than per day. Each is scored and published when it is made, and cannot be withdrawn; the board ranks a team by its best, so a second attempt improves on the first or does not.
• Format checks are unlimited in every phase and spend nothing. They report whether a file is well formed — cell count, genes, coordinates, missing values, structure — and nothing about how it would score.
• A rejected submission does not consume a scored attempt, but every upload counts toward a separate daily upload limit, which exists to bound storage rather than to limit scoring.
• A single prediction file may not exceed 1200 MB.
• Agent Team submissions are held until at least two kinds of evidence are attached — trajectory, prompts, or harness — and only then enter the scoring queue.

12. Evaluation and Ranking

• Every submission is scored by the same published metric panel, on a scale where the floor model sits at 50 and the attainable ceiling at 100.
• A team’s position on a board is its best score on that board. The overall standing scores each task as the mean across its boards, and sums the three, so a board never attempted counts as zero.
• Ranking is within a team’s own track, so each track has its own first, second and third place.
• Final rankings are determined by the hidden test sets, not by the development leaderboard.
• The Organisers’ scoring is final, save for demonstrable error in the evaluation itself.

13. Finalists, Winners and Prizes

Prizes are awarded per track: one first prize, two second prizes and three third prizes in each. Winners are announced at NeurIPS on 11 December 2026.

• Winning teams may be asked for documentation or code sufficient to reproduce the submitted result before a prize is paid. A result that cannot be reproduced is not eligible.
• Agent Team prizes additionally require evidence that the work was the agent’s: the trajectory, the prompts, and the harness. What cannot be verified is not eligible for a prize — it still scores and still ranks.
• Prize recipients are responsible for taxes and for providing the information needed to make payment.

14. Ownership, Licence and Publication

• You keep ownership of your method, your code and your models.
• You grant the Organisers a non-exclusive, royalty-free licence to evaluate your submissions, publish their scores, and describe your method in reports about the Challenge.
• The Organisers may publish aggregate results, leaderboards and technical summaries, respecting the intellectual property position above.
• Winning teams are expected to publish a description of their method, and are invited to contribute to a joint report on the Challenge.

15. Conduct and Disqualification

The Organisers may disqualify an entrant or a team, remove submissions, or withhold a prize, for:

• multiple accounts, or membership of more than one team;
• use of held-out data, or any attempt to recover it;
• misrepresenting how a submission was produced, including a Human Team entry presented as an Agent Team one or the reverse;
• interference with the evaluation service or with other entrants;
• or any other breach of these Rules or of the Terms of Use.

Where a breach is found after a prize is announced, the Organisers may reassign it.

16. Schedule

The schedule is published on the timeline page and forms part of these Rules. Final submissions are due 2 December 2026, official evaluation begins 4 December 2026, and results are announced at NeurIPS.

The Organisers may extend a deadline; a deadline is never brought forward without notice.

17. Compliance, Liability and Governing Law

The Challenge is provided as described in the Terms of Use, including its disclaimers and limitation of liability. These Rules are governed by the laws of the State of California, without regard to conflict-of-law principles.