the evals are the question/answers, ARC-AGI doesn't share the questions and answers for a portion so that models can't be trained on them, the public ones... the public knows the questions so theres a chance they could have been at least partially been trained on the question (if not the actual answer).

Thats how i understand it