Cell area is the joint probability that the LLM assigns to the prefix from the root token to that cell — i.e. P(prompt followed by these tokens). Sibling cells at any subtree sum to that subtree's joint probability; the entire treemap sums to 1.0.
At every internal node, a synthetic (other) leaf carries
the residual mass that wasn't expanded — long-tail tokens dropped by
--top-k / --top-p / --prune-probability
all fold into it. The aggregates therefore reconcile exactly at every
level.
Time flies like an arrow. Fruit flies like a
max-depth / pruned / eos /
other-bucket).