Publication, Corrections & Independence Policy

This Policy explains what VetEval publishes, on what basis, how we handle disputes and corrections, and how we manage the conflicts of interest that come from being operated by a company that also builds veterinary AI. It supplements our Terms of Service and the Methodology.

1. What we publish, and on what basis

We publish measurements: index scores, confidence intervals, rank tiers, category and species breakdowns, safety-gate results, verification tiers, attempt disclosures, and analysis derived from them. Every published number is produced by the method described in the Methodology, on pinned items, harness, and scoring, so a published figure can be tied to exactly what produced it.

Findings are honest evaluation, stated on a disclosed basis. Where we characterise a model's behaviour, including a safety-gate severity, we do so as our assessment of observed model outputs under stated test conditions, on a published methodology, in good faith, and in the public interest of buyers and the veterinary profession. We do not state or imply that a vendor acted wrongfully, and a safety finding describes what a model produced on our items in that run rather than a general claim about the product or the company.

2. Before we publish

• Human review of safety findings. A critical or high safety-gate result is human-confirmed before it can affect a published index.
• Deterministic scoring. Every scored item is multiple choice with a fixed answer key, and no model grades any answer, so a published score can be recomputed from the pinned run record.
• Community review. Community results are checked by our team before they reach the board.
• Notice. We notify the submitting organisation when a result is approved for publication.

3. Permanence, and what we will not do

Community results are permanent. If a model is submitted on a Community plan, the result publishes in full and stays published. We will not remove it because it is unfavourable, because the vendor has withdrawn the model, because a newer version exists, or because the account was closed. Permanence is what makes the board citable; a benchmark you can quietly exit is not a benchmark.

We will not: change a score in exchange for payment or any other consideration; suppress, delay, or de-rank a published result at a vendor's request; remove a result because a vendor threatens litigation, where the result is accurate; or publish a result we know to be wrong.

4. Corrections, annotations, and withdrawal

We distinguish between fixing an error and hiding a result. The table sets out what we do.

Situation

What we do

Public record

Platform or harness error, wrong settings, mis-scoring

Re-run or recompute, correct the figure

Corrected, with a dated correction note

Scoring or item defect found later

Re-grade or retire the item; recompute affected results

Corrected, with a note; item flagged for review

Wrong endpoint or model version evaluated (our error)

Re-run at no cost to the vendor

Superseded entry annotated

Wrong endpoint or version supplied by the vendor

Vendor may resubmit under normal plan limits

Original stands, annotated

Vendor disputes an accurate result

Right of reply published alongside

Result stands

Contamination or manipulation found

Disqualify or annotate the result

Annotated on the record

Legal requirement to remove

Comply, narrowly

Removal noted where lawful

Corrections are dated and visible. We do not silently edit a published number.

5. Your right of reply, and how to raise a dispute

If you are a model provider and believe a published result is inaccurate, was produced under a configuration error, or is materially misleading, contact us through our contact form with the specific figure, the run, and the basis for your concern. Our process:

• We acknowledge receipt and tell you who is handling it.
• We review the pinned run record: items, harness version, scoring, settings, and logs.
• Where the concern is substantiated, we re-run, recompute, or correct, and publish a dated correction note.
• Where the result stands, we tell you why, and you may submit a concise statement of reply, which we publish alongside the result.
• Where we and you continue to disagree, the disagreement itself is noted on the record.

We aim to acknowledge promptly and to resolve substantiated issues quickly. Safety-related concerns are prioritised.

6. Independence and conflicts of interest

We state the conflict plainly rather than claiming to be an unrelated third party. VetEval is operated by viggoVet, which also builds veterinary AI, and it earns revenue from some of the organisations whose models it ranks. Both facts are managed by structural controls and disclosure, not by assertion.

• First-party models are never ranked. viggoVet products do not appear on the public leaderboard. They are validated separately and published as viggoVet research.
• Team separation. Controls keep the benchmark team apart from the model team, as described in the Methodology.
• Data firewall. viggoVet does not train, fine-tune, or otherwise incorporate VetEval evaluation data, held-out items, questions, gold answers, or transcripts into any viggoVet model.
• Platform-enforced integrity. Practice and scored items are separate database states with triggers preventing crossover; scored runs use fixed decode settings enforced by the harness; an endpoint whose data retention was never declared is refused a scored run; every run pins its items, harness, and scoring.
• Governance. Governance parameters are set by VetEval's licensed veterinary reviewers and versioned with the dataset.

7. What payment does and does not buy

Payment can buy

Payment can never buy

Capacity: more models, more evaluations

Rank, or any change to a score

Privacy: seeing your score before publication

Removal of a published Community result

Speed: publishing without waiting for review

Extra attempts on one submission (capped by methodology)

Choice: deciding whether a Lab result publishes

Secrecy about how many attempts preceded a result

Standard plan pricing is uniform and published to every customer, and we do not negotiate the price of a standard plan, because we rank the organisations that pay us.

8. Use of third-party names and marks

We identify models, vendors, and their marks nominatively, only to say what was evaluated, using no more of a mark than needed and never in a way that suggests affiliation, sponsorship, or endorsement in either direction.

9. Changes

We may update this Policy and will publish the updated version with a new date. Material changes to how we publish, correct, or disclose conflicts will be described in the changelog.