What the model actually said.

Ten example trajectories per task, straight from the benchmark: the exact stimulus, the answer choices, the ground truth, and the model's full response.