The components are AA-Omniscience-Decision, HLE-MultiChoice, When2Call, and WildGuardTest, covering knowledge and hallucination, reasoning, tool calling, and classification.
The page says AA-Omniscience-Decision uses 500 private questions with four candidate answers plus “I don't know”; accuracy and non-hallucination rate each contribute 12.5 percentage points to the index. HLE-MultiChoice uses 513 text-only multiple-choice questions from Humanity's Last Exam. It also distinguishes the response handling for decision models from language models, including structured outputs for the latter.
The disclosure makes the aggregate index easier to inspect and interpret, but does not establish that its task mix represents every deployment context. Teams can use it as a shortlist and should validate leading candidates against their own labels, traffic, and error costs.