Odyssey Launches CaliBench for World Model Randomness

Odyssey posted an announcement introducing CaliBench. The benchmark evaluates whether video world models reproduce the true randomness of physical events such as dice rolls or card draws. A linked page on the company site describes the test as checking if models produce outcome distributions that match reality rather than only appearing physically plausible frame by frame. Visible replies include one noting a generated video where an apple floated instead of falling and another observing that models show biases toward certain outcomes.