Maintained by OneHill. Status: v0.1, Foundation. Content baseline: July 2026.
Adopting an open model, or routing your traffic through a hosted inference API, is an ownership decision as much as a capability one. Four things decide how much of it is actually yours:
- Use and modify it freely, with no gate and no field-of-use trap.
- See what it is: the weights, the training, the behaviour, and the binding terms.
- Rely on it: dependable to run and safe to deploy. How capable it is shapes the AOI score, not this factor - you can fully own a model that is not the most capable.
- Keep your data yours: running it does not hand your inputs or your knowledge to someone else.
The AI Ownership Index rates every entry on those four factors and resolves them into one
verdict: an ownership level, from full down through substantial, partial and
limited to none. The verdict is floor-weighted, so the weakest factor caps it. A black
box is never "full", however cheap the price.
The AOI score is a 0 to 100 headline built from seven weighted dimensions, each scored 0 to 5 against a published anchor. It is an analytical input to the ownership verdict, not the verdict itself. The rubric is public and versioned; see methodology/scoring-rubric.md and methodology/versioning.md.
Models are scored on Openness and Transparency, Provenance and Supply-chain Integrity, Legal and Regulatory Readiness, Safety and Alignment, Technical Performance, Operational Readiness, and Maintenance and Governance. Inference providers use a parallel set: data governance, compliance, residency, security, reliability, transparency and lock-in, and cost. Every rung 0 to 5 is defined for all fourteen dimensions; the exact anchors and the ownership cutoffs are laid out in the classification matrices.
| Section | What it documents |
|---|---|
models/ |
One folder per open model: a validated entry.yaml plus a four-domain dossier (assess, implement, use, support). |
inference-providers/ |
APIs that serve open models: retention, training-on-inputs, residency, compliance, and security, read from the binding terms. |
methodology/ |
The versioned rubric, the ownership rules, the openness framework, the EU AI Act mapping, and the grounding standard. |
schema/ |
JSON Schemas that every entry validates against. |
scripts/ |
Scoring, completeness, and validation tooling. Scores are derived from the anchors, never hand-typed. |
templates/ |
Copy-to-start folders for contributing a new entry. |
site/ |
The published site: the library ledger, the entry sheets, compare, and the reference pages. |
Every rating traces to a primary document, tagged with which kind it is (doc_type) and
whether the text was actually read (retrieved). A binding document, a licence, a signed
DPA, or read terms, along with an independent attestation, outranks a claim on a product
page. A dimension scored 5, and any transparency or data-control factor rated strong, must
cite a retrieved binding document or a third-party attestation; a self-reported claim is
marked, dated, and cannot reach the top anchor. CI enforces this on every change. Where a
fact may already have moved, a version, a price, a certification, the entry says so and
dates it.
The reader-facing index is at ownershipindex.ai: a ledger of every entry ranked by AOI, an entry sheet per model and provider that leads with the ownership verdict, side-by-side compare, and reference pages for the classification matrices and the glossary. The site is a projection of this repository, so what a reader sees and what they can audit are the same thing.
New entries and corrections are welcome by pull request. The rules, the evidence-first score-change process, and the review requirements are in CONTRIBUTING.md; how the project is run is in GOVERNANCE.md; the prose rules (plain hyphens, British spelling, no filler) are in STYLE.md.
No index can independently re-run every evaluation on every open checkpoint. Where a datum
comes from a publisher or a third party rather than from OneHill's own check, it is marked
as such (source_type: publisher | third_party | onehill_verified). A missing document is
recorded as a finding, not papered over. Trust the labels.
This repository is licensed under the Apache License 2.0. See LICENSE and
NOTICE. Primary-source extracts under docs/primary-sources/ remain the property
of their respective publishers.