This experiment begins with a short prompt and generates a main sequence. Each token in that sequence then becomes the starting point for a separate continuation, with the surrounding context removed.
The token is decoded to text and encoded again as a standalone prompt. That can change its tokenization, so the experiment records both forms.
The output is intended to make the continuations comparable in a tree view and other visualizations. The question is descriptive: what directions does the model take when it receives a fragment of an earlier output without the context that produced it? The experiment does not assume those branches reveal a hidden reasoning process.