Glossary
Plain definitions of the terms used on the site.
Scores and verdicts
Section titled “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 closing level for how much you own the model or provider. For a model it concerns the weights (can you run, adapt and specialise them). For a provider it concerns your data. The levels are full, substantial, partial, limited and none.
- Completeness - the share of the expected documentation items that an entry covers. Below 100% means some items are gaps. It measures coverage and says nothing about quality.
Provenance badges
Section titled “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. No entry carries one yet.
- gap - something that could not be gathered or provided, recorded with the reason.
Openness
Section titled “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
Section titled “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
Section titled “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
Section titled “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.