Two roads diverged. Pick one.
Direction surfaces a latent structure in language models: every next token has a distribution, and two high-probability draws can be radically different paragraphs. The left/right choice UI externalizes what normally happens invisibly — the model's uncertainty at a fork. Post-trained models produce more interesting forks than base models: their RLHF tuning creates sharper stylistic modes, so Option A and Option B feel like genuine perspectives, not random variation. The "down to regenerate" action is a direct sample from the same distribution — every press reminds you that the model's imagination is not a single road but a branching tree, and you are its navigator.