See how the model's next-token probability distribution changes across training stages for a fixed prompt. Toggle between stages to see which continuations become more or less likely.