Why trust this index

You are right to distrust ratings you cannot check. This page explains why this index is different: every rating traces to a document you can read, the method is public, and the score is built to recognised standards rather than to opinion.

The one rule everything rests on

Every claim in an entry must trace to a primary document that was actually read, not a marketing page. Each piece of evidence carries two auditable fields: what kind of document it is, and whether its text was actually read (retrieved). A strong rating cannot rest on a marketing page or an unread document. Where a document is missing, we record that as a finding rather than leave it unstated. You can see this on any entry: the Sources section lists each document, whether it was read, and a link.

This is also why the Runware entry says its data terms are weak: we read the binding Terms and Privacy Policy, and they contradicted the marketing. We report the document, not the brochure.

Independence

The Index is stewarded by the OneHill Foundation. Ratings are not for sale and entries are not pay-to-play. Where a provider or model publisher disputes a rating, they are invited to point at the primary document that supports their position, and the entry is updated from the read text. See GOVERNANCE.

Standards it is built to

The scoring is grounded on public standards and frameworks, linked below. We follow the same rule we apply to entries: link the canonical source, and keep a short dated extract where a standard defines something we use, rather than copying the whole document.

• Model Openness Framework - the openness-tier taxonomy (fully open vs open weights) our tiers mirror.
• OSI Open Source AI Definition - the bar for what counts as open source AI (usable data information, code, and weights).
• Stanford Foundation Model Transparency Index - transparency indicators behind the openness and provenance dimensions.
• EU AI Act (Regulation (EU) 2024/1689) - Articles 53 and 55 duties and the systemic-risk compute threshold, shown on each model’s EU AI Act fact.
• NIST AI Risk Management Framework - the risk framing behind the safety and governance dimensions.
• OWASP Top 10 for LLM Applications - the deployment risks behind the safe-deployment and security guidance.
• MITRE ATLAS - the adversarial-ML threat taxonomy behind supply-chain and safety.
• OpenSSF and SLSA - supply-chain integrity and build-provenance behind the provenance dimension and format-safety checks.

Keeping it current

Each entry carries a last_verified date and a freshness target. A weekly review ingests newly gathered primary documents, re-grounds the affected entries, re-runs validation, and reports what changed and what is still outstanding. The data is validated in continuous integration on every change, so an entry that does not meet the grounding rule does not ship.

ownershipindex.ai and GitHub

The reader-facing site is published at ownershipindex.ai (the AI Ownership Index). The full source, the calculations, the methodology, and the primary-source extracts live in the public GitHub repository, OneHillAI/AOI, which anyone can inspect or contribute to. The published site is a projection of that repository, so what you read and what you can audit are the same thing.

Questions or corrections: dev@onehill.org.