Governance Charter

This Governance Charter establishes the principles, processes, and accountability structures for the operation of VAULT (Veterinary AI Unified Leaderboard & Testing), a benchmark platform operated by ANIML Health.

VAULT exists to provide the veterinary AI community with a rigorous, transparent, and trustworthy evaluation standard for clinical AI tools. This Charter governs how the benchmark is designed, how access is controlled, how results are evaluated, and how the leaderboard is maintained.

VAULT is operated exclusively by ANIML Health. A designated Benchmark Governance Committee (BGC) consisting of ANIML Health staff oversees all material decisions regarding the benchmark, including:

• — Benchmark dataset curation and anonymization protocols
• — Grader version releases and deprecation
• — Participant access approvals and revocations
• — Leaderboard publication approvals
• — Material amendments to this Charter or related policies

The VAULT benchmark dataset comprises 5,000 anonymized veterinary clinical records curated from real clinical encounters. The dataset is the intellectual property of ANIML Health and subject to the following guarantees:

• — All records are de-identified in compliance with HIPAA-equivalent standards
• — The dataset undergoes internal IRB-equivalent review before use
• — Records are never exposed to benchmark participants in readable form
• — No raw record or reconstruction-enabling subset is returned in any report or API response
• — The dataset is stored encrypted at rest and in transit

The VAULT grader is a proprietary internal evaluation system. To protect benchmark integrity:

• — Participants cannot submit, substitute, or inspect the grader implementation
• — Grader rubric summaries are publicly documented; detailed rubrics are internal
• — All grader versions are immutably versioned and logged
• — Scores from different grader versions are not directly compared on the leaderboard without explicit notation
• — ANIML Health reserves the right to re-evaluate past runs if a grader defect is discovered

Publication of results to the public leaderboard requires two conditions:

(a) Participant consent — Participants must explicitly consent to publication. Results are private by default.

(b) Admin review — The Benchmark Governance Committee must approve the submission. Approval may be withheld if results are suspected to be the product of gaming, methodological violations, or technical error.

Published results include: model name, organization, benchmark suite version, composite and sub-scores, median latency, and date. Raw outputs, per-case scores, and reconstruction-enabling data are never published.

ANIML Health staff involved in the benchmark evaluation process are prohibited from submitting their own models for evaluation under the same review process. Internal evaluations must be disclosed as such on the leaderboard.

Participants who believe a score is erroneous may submit a formal dispute by email to benchmark@animl.health within 30 days of run completion. The BGC will review disputed runs within 10 business days. Decisions are final.

ANIML Health may amend this Charter at any time. Material amendments will be communicated to active participants via email at least 14 days before taking effect. Continued participation after the effective date constitutes acceptance.