Skip to content

Methodology

Every rating here is an opinion with receipts. The full method lives in the GitHub repository; this page summarises it.

Each model is scored on seven weighted dimensions, each 0-5, rolled into a 0-100 headline and an A-F grade. Trust dimensions deliberately outweigh raw capability - a model that benchmarks well but can’t be trusted, licensed, or run safely is the failure this library exists to catch.

Dimension Weight
Openness & Transparency 0.18
Provenance & Supply-chain Integrity 0.16
Legal & Regulatory Readiness 0.16
Safety & Alignment 0.16
Technical Performance 0.14
Operational Readiness 0.12
Maintenance & Governance 0.08

Hard flags (e.g. a systemic-risk model with no EU AI Act Article 55 documentation) can cap a grade regardless of the computed score.

Beyond the trust score, each documentation domain shows:

  • Completeness - coverage of the expected items for that domain (grounded in accepted standards). A documented gap counts as addressed-but-empty, not as coverage.
  • OneHill-tested (originality) - the share of a domain’s items OneHill generated or verified itself, rather than aggregating from the publisher’s own docs.

Scoring and risk: the Model Openness Framework, OSI’s Open Source AI Definition, Stanford’s Foundation Model Transparency Index, the EU AI Act (GPAI obligations), NIST AI RMF, OWASP LLM Top 10, MITRE ATLAS, and OpenSSF model-signing. Documentation structure: Diátaxis, ISO/IEC/IEEE 26514, and the Good Docs Project. See Documentation taxonomy and How we source it.