Definition
Plain language
A published report describing what a new AI model can do and what it got wrong in testing.
As stated in the literature
A release document covering capability evaluations, safety testing, red-teaming results, and mitigation decisions for a specific model version; broader in scope than a model card.
Also called: System Card, system cards
Why it matters: It is often the main public window into what a model can really do, so its thoroughness shapes what regulators, researchers, and users can even discuss.
For example, before releasing a new model a lab might publish a document describing how it performed on hacking tests, where red-teamers got it to misbehave, and which safeguards were added as a result.
Heard on the show
“Frontier labs now routinely report evaluation awareness in their system cards — models detecting when they're being tested at rates above chance — and the capability scales with model size.”Episode 128 — How a Model Can Earn Full Reward and Still Resist Training