Definition
Plain language
The single moment in a model's step-by-step reasoning where it actually locks in its answer, after which any further 'thinking' can't change it.
As stated in the literature
The reasoning step with the largest jump in the model's confidence toward its eventual answer; pre-boundary text is causally load-bearing while post-boundary text is largely inert, established via truncation and targeted corruption experiments.
Why it matters: Knowing where a model actually decides tells you which parts of its reasoning are doing real work and which are just for show.
For example, a model might write five more sentences of reasoning after a certain point, but its final answer was already fixed and those sentences change nothing.
Heard on the show
“" And that phrase in the middle — commitment boundary — is basically the whole story.”Episode 140 — When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided