The open verification process is convenient, though I would prefer more details on the submission workflow for smaller teams.
Questions about EvalSeal
What is EvalSeal?
EvalSeal issues signed, tamper-evident records for AI evaluation runs that capture dataset hashes, model build information, sampling conditions, and traces using Ed25519 cryptographic signing. The service acts as an independent record-keeping system that does not publish scores or endorse results. These records allow third parties to audit, contest, and overturn evaluation findings without needing to trust EvalSeal itself.
How does EvalSeal handle scoring and endorsements?
The company explicitly states it does not publish scores or endorse results, focusing instead on issuing receipts for evaluation runs. This approach addresses common problems with AI capability claims such as data contamination, irreproducibility, and model drift. By freezing a single run into a record, they ensure the evidence remains unaltered and verifiable by outsiders.
What specific details are captured in an EvalSeal record?
Each record includes a dataset hash, model build, sampling conditions, and every trace, all signed to prevent tampering. The task set is frozen and hashed before the run, meaning swapping or editing a single task stops the hash from matching. This ensures the test set cannot be altered after the evaluation begins.
How does the hash mechanism prevent tampering?
The website states that the task set is frozen and hashed before the run, with the hash stopping if even a single task is swapped or edited. This mechanism ensures that the evaluation conditions remain exactly as recorded. The system uses Ed25519 cryptographic signing to secure the integrity of the entire record.
Who is the ideal user for EvalSeal services?
The service is designed for anyone needing to audit, contest, or overturn AI evaluation findings. It specifically addresses issues like data contamination, irreproducibility, and model drift in AI capability claims. The independent nature of the records allows third parties to verify results without relying on the vendor's own assessment.
What regulatory obligations does EvalSeal address?
The company's own website mentions that the AI Office can levy fines against providers of general-purpose models starting from August 2025. Annex XI requires providers of systemic-risk models to document evaluation strategies including results and adversarial testing. Evaluators must be able to evidence their own independence when findings are passed to regulators.
What are the current regulatory timelines and requirements?
The website notes that obligations for high-risk systems are being pushed to December 2027. The AI Office has been active since August 2026, and the machinery to enforce these obligations is now in place. Providers must keep files for ten years, and external evaluators must document their independence.
What does the public register currently show?
The website displays a public register showing one seal issued and zero challenges filed as of the last update on 2026-07-28. The register allows users to open specific seal numbers in full to inspect the details. This transparency is part of the independent record-keeping system EvalSeal promotes.
Where can users find EvalSeal's terms and privacy policies?
The website includes a footer link to 'Terms & privacy' which outlines the business's policies. The company also provides a link to 'Write to us' for general inquiries. These sections are part of the standard business profile information available on their site.
How does EvalSeal ensure accurate timestamps?
The website mentions using time anchors by freetsa.org to ensure the records are anchored in time. This feature helps prevent backdating or manipulation of the evaluation timestamps. The use of external time anchors adds another layer of verification to the records.
What specific evaluation problems does EvalSeal solve?
The company's own website states that the default evaluation process does not produce evidence due to issues like contamination, irreproducibility, drift, and single-party bias. EvalSeal's records are designed to solve these specific structural problems in AI evaluation. The service focuses on creating evidence that can be refuted, which is a core requirement for valid evaluation.
What is the official website of EvalSeal?
The official website of EvalSeal is evalseal.com.
How is EvalSeal rated on Trustburn?
On Trustburn, EvalSeal has a rating of 4.2 out of 5 based on 5 reviews.
Post your question publicly — other visitors, and whoever manages this page, can answer. It will be visible on this page.
Want to hear about this company again? Follow it and we’ll email you when a new review appears.
Based on reviews, listing data and the company's website. Verify anything important with the company directly.
The approach to handling AI evaluation data seems structurally sound. By focusing on issuing a receipt with a frozen hash rather than a subjective verdict, EvalSeal addresses the core issues of contamination, irreproducibility, and drift. The text explains that scores often climb due to test set leakage, but a frozen hash prevents this if the task set remains unchanged. The distinction between grades A, B, and C is particularly helpful for understanding the level of independence. Grade A ensures the subject had no access to the environment, while Grade C only attests to the integrity of submitted traces. This clarity is vital for regulatory compliance, especially with the AI Office requiring documentation for general-purpose models and high-risk systems. The fact that the record is fixed at the moment of the run and any edit is arithmetically obvious adds a layer of trust that is often missing in standard evaluation reports. It binds the provider and evaluator in a verifiable way, supporting the ability to audit and contest results effectively.
The clear explanation of how a Grade A seal differs from a Grade C one is exactly what was missing. It removes the guesswork about who actually executed the run and how the environment was provisioned.
Reading through the available information, EvalSeal positions itself as an issuer of receipts rather than verdicts for AI evaluations. The distinction between grades A, B, and C is critical; Grade A involves EvalSeal provisioning the environment and executing the run independently, whereas Grade C merely attests to the integrity and time of submitted traces without confirming the run happened as described. The text highlights issues like contamination, irreproducibility, and drift in standard evaluations. It notes that the AI Office can levy fines for general-purpose models, requiring documentation kept for ten years. The service seems to address the need for auditable, frozen records with specific dataset hashes and model build details, ensuring that any later edits are arithmetically obvious, which supports the goal of making evaluations contestable and verifiable by third parties.
The frozen hash concept is genuinely useful for tracking model drift, though I wish they clarified exactly which AI Office regulations apply to smaller firms.
Questions & Answers
No questions yet — be the first to ask.
Sign in to post
Your text is saved — you'll come right back to post.
Don't have an account? Create one
News & Press
No news yet. This is where the business shares its own updates and press.
Photos
No photos yet. Photos shared by reviewers will show up here.
About EvalSeal
EvalSeal provides a service that issues signed, tamper-evident records for AI evaluation runs, capturing details such as dataset hashes, model build information, sampling conditions, and traces. These records use Ed25519 cryptographic signing and are designed to make evaluation results auditable, contestable, and reproducible by third parties without requiring trust in EvalSeal itself. The company explicitly states that it does not publish scores or endorse results, but rather serves as an independent record-keeping system for AI evaluations. The service is positioned as a way to address common problems with AI capability claims, including data contamination, irreproducibility, model drift, and conflicts of interest when vendors evaluate their own systems.
- Website
- evalseal.com
