The numeric check is the one that matters
Two answers that read similarly and quote different figures is the failure worth catching. Prose differences are usually stylistic; a differing number means one of them is wrong.
Numbers present in one answer and absent from the other are listed separately for that reason, ahead of everything else in the report.
What overlap does and does not tell you
This measures shared vocabulary, not shared meaning. Two correct answers phrased differently score low; a fluent wrong answer reusing the same words scores high.
Use it to find what to read, not to decide which answer is better. For that, an LLM judge with an explicit rubric is the appropriate instrument.