The pelican has looked very same-y across all frontier models, same color bike, same camera angle, etc. I suspect this challenge is already too embedded in the training data to be a good signal when it succeeds, and maybe even when it fails in pathological ways mirroring existing AI pelicans on the internet.