Models with similar averages can fail in completely different ways. Report Cards automatically write model behavior into reports — and verify the reports themselves.
Launch Highlights
- Problem: A single average score hides where a model succeeds, fails, and changes behavior.
- Method: Report Cards generate natural-language behavior summaries and evaluate them with contrastive, Elo, and human scoring.
- Finding: Strong reports compress many examples into evidence that helps people tell models apart.
- Why it matters: Evaluation results can feed product review, model choice, and safe deployment, rather than stopping at a leaderboard.
Continue reading
The research page covers the background, method, key figures, and paper links. Enterprise, government and research teams: contact the lab for evaluation details or research collaboration.