Official Competition Rules

Version 2026-08-17 · Agenthon 2026 — NeurIPS 2026 Competition Track · AH26-POL-01
Policy set: Official Competition Rules · Terms of Participation · Privacy Notice · Data & Software Licensing Policy

These Rules govern eligibility, teams, submissions, permitted resources, evaluation, integrity, final rankings, reproducibility, prizes, and the NeurIPS final event. Track instructions may add technical requirements but may not silently change participant intellectual-property or privacy rights.

Operational principle. A submission is ranked only after it passes the applicable admissibility gates and is successfully evaluated under the published protocol.

1. Scope and acceptance

Agenthon 2026 (the “Competition”) is a research competition in the NeurIPS 2026 Competition Track focused on verifiable AI for quantitative finance. By registering, accessing the competition environment, or submitting, each participant agrees to these Rules, the Terms of Participation, the Privacy Notice, the Data & Software Licensing Policy, applicable track instructions, and relevant platform terms.

Participants are responsible for reading official notices and ensuring that their team, submission, resources, and conduct remain compliant throughout the Competition.

2. Organizers and official channels

The Competition is organized by the Society of Quantitative Analysts (SQA) and CEWIT at Stony Brook University (together, the “Organizers”), subject to applicable institutional approvals. Track leads, evaluators, Question Partners, sponsors, and platform or compute providers may support operations without acquiring ownership of participant submissions merely because of that role.

• Official site: agenthon.net.
• Official notices: posts on the Competition Site or platform, and emails sent to the registration address.
• Rule and technical questions: the designated forum or admin@agenthon.net. Written public announcements control over informal statements.

3. Competition schedule

The Competition Site and platform clock are authoritative. NeurIPS 2026 runs 9–13 December 2026 in Atlanta, Georgia; the Agenthon session falls within those dates, most likely on one of the final days, and the exact day is designated by NeurIPS. Site-specific schedules and later NeurIPS instructions control. Material schedule changes will be announced through an official channel and, where practicable, applied prospectively.

4. Eligibility, teams, and accounts

• Participants must be at least 18 and have reached the age of majority in their jurisdiction. Participation and prize payment must be lawful and may be subject to sanctions, export-control, anti-bribery, tax, and other legal screening.
• Organizing Committee members, designated judges/evaluators, persons with material access to non-public test materials, and members of their immediate household are not eligible for prizes. Sponsor or Question Partner personnel may participate only without privileged access and subject to track-specific conflict rules.
• Participants must obtain any employer, university, sponsor, or other approval needed to participate or submit work created in the course of employment or study.
• Teams may have one to three registered participants. Each person may use one competition identity and belong to one team only.
• Each team must designate a captain for official communications and final-submission decisions. Credentials may not be shared outside the team.
• Team changes require Organizer approval and ordinarily must be completed before private final evaluation. Changes may not transfer non-public solutions or circumvent submission limits.

5. Tracks and competition structure

Teams may enter one or more tracks. Track instructions form part of these Rules and may specify interfaces, schemas, resource limits, data cutoffs, scoring formulas, tie-breakers, and submission procedures.

6. Submission and resource requirements

• Official submissions must use the approved Docker or other containerized format and implement the stable command-line interface for the entered track.
• Final submissions must run without human intervention in the designated sandbox. External network access should be assumed unavailable unless a track expressly permits it.
• Submissions must satisfy published size, runtime, memory, dependency, security, and interface limits. Unsafe, unreliable, or non-executable submissions may be rejected as inadmissible.
• Submissions must not contain malware, destructive code, credential harvesting, unauthorized scanning, hidden communication channels, or functionality designed to interfere with infrastructure or third parties.
• Do not embed personal data, private credentials, API keys, confidential employer information, or data that the team is not authorized to provide.
• Unless a track states otherwise, each team designates one final submission per entered track.

6.1 External resources and provenance

• Pre-existing software, public data, pretrained models, research code, and commercial tools may be used only when lawful, properly licensed, and permitted by the track.
• Participants remain responsible for material produced with AI agents, language models, code assistants, or other automated tools.
• Prize-eligible teams must provide the resource/provenance declaration required by the Data & Software Licensing Policy.
• Private, employer-only, sponsor-only, embargoed, or other non-public data may not be used unless expressly authorized and supported by documented permission.
• A track cutoff or as-of date applies to features, labels, prompts, retrieval indexes, caches, fine-tuned weights, synthetic data, and other derived artifacts — not only to raw inputs.

7. Public/private firewall, admissibility, and ranking

• Each track may separate public practice resources from a sealed private evaluation environment. Private-test units, hidden labels, oracle answers, final scoring logic, canary registries, audit materials, and equivalent assets are confidential Competition materials.
• Participants may not access, infer, reconstruct, scrape, exfiltrate, retain, or share non-public test material outside the normal scoring interface.
• Leaderboard probing, adaptive query patterns, reverse engineering of hidden labels, exploitation of scoring bugs, and other techniques intended to extract private-test information are prohibited.
• Official runs may pass through g0 integrity, g1 schema/interface, g2 cutoff/resource, and g3 domain-semantics checks. Only admissible runs receive the applicable metric.
• Development leaderboard scores are provisional. Final rankings use designated private data and the final scoring implementation, subject to verification and correction of demonstrable evaluation or infrastructure errors.
• Published track tie-breakers control. If no tie-breaker is published and scores are exactly tied at the designated precision, the teams remain tied.
• A team may request review of a demonstrable scoring, execution, or infrastructure error within the period stated in the results notice. Review does not permit a new submission or a post-deadline change to the metric.

8. Private verification and winner public release

Agenthon uses a three-stage model: ordinary submissions are private; provisional prize candidates undergo confidential verification; and only a confirmed winner that elects to accept a prize must publish a separate reproducibility package.

• A provisional prize candidate must provide participant-authored source code, build/run instructions, dependency versions, seeds/configuration, model or artifact information, a provenance declaration, and a concise method summary upon request.
• The Organizers may rerun the submission with fresh seeds or equivalent controls and inspect code, container contents, logs, manifests, and declarations only as reasonably necessary for reproducibility, integrity, eligibility, and security.
• A result that cannot be reproduced to reasonable tolerance may be corrected, removed, or deemed ineligible for a prize.
• After written confirmation of winning status, a team accepting the prize must publish the participant-authored code needed to reproduce the winning method under an OSI-approved license, ordinarily within 14 days.
• The public package may be a cleaned repository and need not be the exact private container. Teams need not disclose secrets or relicense third-party data, software, models, or weights they do not own, but must document a lawful reproduction path.
• Declining or failing the public-release condition may forfeit the prize without transferring ownership. Unless the result itself is invalid, the scientific ranking need not change solely because a winner declines the prize.

9. Integrity, conduct, and enforcement

Participants must comply with the NeurIPS Code of Conduct and applicable NeurIPS policies concerning professional conduct, academic integrity, anti-collusion, and responsible use of agents or language models to the extent those policies apply.

Prohibited conduct includes plagiarism or misappropriation; collusion or private sharing of non-public solutions; multiple accounts or false identities; unauthorized access to test materials, scoring logic, other submissions, or restricted systems; attacks on the platform or infrastructure; falsified declarations or results; and score manipulation through bugs, prompt injection, or hidden side channels.

The Organizers may investigate and impose proportionate measures, including warning, score removal, submission rejection, suspension, disqualification, prize forfeiture, or removal from official results. Where feasible and consistent with integrity or security needs, the affected team will receive the material reason and an opportunity to respond before a final prize-related decision.

Good-faith reporting of a benchmark, privacy, or security issue will not itself be treated as misconduct.

10. Prizes, publication, and scientific reporting

• Prize categories, amounts, sponsor-funded awards, and special eligibility conditions will be announced through an official channel before they apply to the Final Phase.
• Potential winners must satisfy identity, eligibility, authorship, private verification, tax, and payment-documentation requirements. Unless announced otherwise, a team prize is divided equally among registered members.
• The Organizers may publish team names, scores, ranks, track breakdowns, finalist/winner names and affiliations, approved public method summaries, prize information, and aggregate benchmark analyses.
• Non-public source code, containers, model artifacts, and private verification materials do not become public merely because a result or winning status is announced.
• Publication of a participant-authored manuscript or non-public technical report requires the participant’s authorization.

11. NeurIPS final event

The Agenthon in-person presentation/workshop is conducted as part of the NeurIPS 2026 Competition Track. Participants must satisfy the registration, credentialing, admission, Code of Conduct, safety, security, accessibility, and venue requirements applicable to the assigned NeurIPS site.

• Finalist or winner status does not itself provide NeurIPS registration, a badge, venue access, a visa, travel authorization, or entry into the host country.
• Unless separately announced, participants are responsible for registration, travel, immigration, and venue-entry requirements.
• Alternative presenters or arrangements for exceptional circumstances require written Organizer approval and must be permitted by NeurIPS. No remote or substitute arrangement is guaranteed.
• Third-party decisions concerning registration, credentials, immigration, security, or physical access do not by themselves invalidate a scientific result, although failure to satisfy an expressly announced presentation or prize condition may affect presentation or prize eligibility.

Official references: the NeurIPS 2026 Competition Track and the NeurIPS Code of Conduct.

12. Changes, cancellation, precedence, and contact

The Organizers may clarify rules, correct benchmark defects, extend deadlines, or rerun affected evaluations when reasonably necessary for fairness, security, legal compliance, or scientific validity. Material changes will be announced and, where practicable, applied prospectively.

A material outage, security compromise, legal requirement, force majeure event, or other circumstance may require suspension, modification, or cancellation of an affected phase or the Competition.

For operational matters, a specific track instruction controls over a general Rule. The Terms of Participation control intellectual property, confidentiality, disclaimers, and governing law; the Data & Software Licensing Policy controls resource licenses; the Privacy Notice controls personal-data processing; and mandatory law controls all documents. Written interpretations may resolve ambiguity but may not retroactively create a material new obligation after the relevant deadline.

General competition and policy questions: admin@agenthon.net. Additional published contacts: admin@sqa-us.org and info@cewit.org.