Separate possibility from evidence
We distinguish a capability area from a measured result. Illustrative tasks are labeled, research perspectives are identified as such, and unpublished specifications remain explicitly unpublished.
Keep people in the loop
Model responses can be incomplete or incorrect. When an output affects a consequential decision, human review and suitable domain expertise are essential. Generated code should be inspected and tested before use.
Treat inputs with care
Only use data that you are authorized to provide. Avoid including confidential, personal, or sensitive information in evaluation material without an appropriate basis and handling process. This research website does not offer an input upload or inference service.
Study limits as well as successes
Failures, ambiguity, and contradictory evidence belong in an evaluation. Research should ask when a system should decline, seek clarification, or communicate uncertainty.
Report a concern
For questions about content or concerns about a model description, use the contact page. Include the page, the statement in question, and the context needed to understand the issue.