What this checklist does
The checklist helps an author organize release documentation for a causal, instruction-tuned, or generic open-weight model. It does not run the model, inspect its weights, verify every factual claim, or certify that a release is safe.
Status rules
Documented
A normal descriptive field is non-empty, or an evidence-based field contains the required method, date, result, or source.
Needs evidence
A claim is present, but the information required to make it reviewable is incomplete. For example, “we tested jailbreaks” without a tool, date, and result summary.
Missing
A core release field or evidence record has not been documented.
Evidence requirements in v1.0
| Evidence field | Minimum information |
|---|---|
| Performance benchmark | Benchmark name and score or result |
| Jailbreak evaluation | Tool or method, date, and result summary |
| Harmful-content evaluation | Tool or dataset, date, and result summary |
| Bias evaluation | Method or dataset, date, and result summary |
| Training-data license | License name and source URL |
Reference sources
The checklist draws on public documentation practices and risk-management guidance. These references are sources, not endorsements or certifications.
- Mitchell et al. (2019), Model Cards for Model Reporting
- Hugging Face Model Cards documentation
- NIST AI Risk Management Framework
- OWASP guidance for LLM applications
- European Commission AI regulatory framework overview
Versioning
Future changes will increment the checklist and schema versions. A JSON export records both versions so a future importer can explain what changed before upgrading a saved project.