The file format determines whether loading a model can run code on the host machine. Safer formats only contain numbers; risky formats can execute arbitrary instructions.

The model card is the publisher's primary documentation surface and is referenced directly in EU AI Act Article 13. Each section is checked for real content, not placeholder headings.

An unclear or missing licence creates legal risk for anyone using the model commercially. We score whether a licence is declared and whether its full terms are accessible.

Account history and verification status indicate whether the publisher is who they claim to be. A reused or freshly created namespace is a documented route for distributing tampered models.

We check whether the publisher documented safety and bias testing. The model itself is never run by Dokima; this dimension scores the presence of structured evidence, not its outcomes.

The EU AI Act requires AI systems offered in Europe to publish specific transparency information. We check for the five categories Article 13 names.

A trustworthy model usually leaves a trail beyond the model registry. We look for community activity, external references such as GitHub or research papers, and any third party security flags.