Glossary

Plain-language definitions. If a term on any page is unclear, it should be here.

Scores and verdicts

• AI Ownership Index (AOI) - this index’s 0-to-100 score, with a letter grade, for how open, well-governed and trustworthy a model or provider is. Higher is better. It is built from seven weighted dimensions.
• Grade (A to F) - the band of the AOI score: A is 85 or above, B is 70 to 84, C is 55 to 69, D is 40 to 54, F is below 40.
• Ownership verdict - the single closing judgement on how much you really own the thing: for a model it is about the weights (can you run, adapt and specialise it); for a provider it is about your data. Levels are full, substantial, partial, limited, none.
• Posture - the short strip of letters on the index that flags the handful of facts that most affect a decision. Green is strong, amber is partial, red is weak or absent. Hover a letter for its meaning.
• Completeness - how much of the expected documentation we have actually written for an entry. Below 100% is normal and means some items are still gaps or not yet gathered. It measures coverage, not quality.
• OneHill-tested - the share of an entry that OneHill verified or produced itself, rather than citing from a source. A low number is honest, not a fault.

Provenance badges

• aggregated - a claim taken from an existing source, cited and dated.
• OneHill-tested - a claim OneHill checked or reproduced, with the method recorded.
• gap - something that cannot be gathered or provided, recorded explicitly with the reason rather than quietly left out.

Openness

• Open weights - the trained model file is published, so you can download and run it yourself. The training data and code may still be closed, so you cannot fully reproduce it.
• Fully open - the weights, the training data, and the training code are all released, so the model can be reproduced and audited end to end.
• Licence - the legal terms you accept to use the weights. Some are permissive (MIT, Apache) with no user caps; others attach conditions.

Data and provider terms

• Zero data retention (ZDR) - the provider does not store your prompts or the model’s outputs.
• DPA (Data Processing Agreement) - the binding contract that governs how a provider handles your personal data.
• Sub-processor - a third party the provider uses to run the service, which therefore may touch your data.
• CLOUD Act - a US law that can compel US-based companies to hand over data. It is why a provider’s country of incorporation matters for sovereignty.
• Retrieved - on a piece of evidence, “true” means the actual binding document was read, not just its marketing summary.

Regulation

• GPAI (General-Purpose AI) - the EU AI Act’s category for broadly capable models, which carries transparency duties.
• EU AI Act Article 53 - the baseline duties for GPAI models, with a lighter path for open-source releases.
• EU AI Act Article 55 and the 1e25 FLOPs threshold - extra duties for the largest models, triggered when training compute crosses roughly 10^25 floating-point operations.

Technical terms you may meet

• safetensors - a model file format that cannot execute code when loaded, unlike the older pickle format. Safer to download and run.
• Quantisation - shrinking a model (for example to run on smaller hardware) by storing its numbers at lower precision.
• SLSA - a standard for tamper-evidence in how software is built and released.
• METR - an independent organisation that evaluates AI models for dangerous capabilities.