It was a weird bug (which I think is still present in the codebase) that counted the original words wrong because of a non-breaking space or something. The counts are not really exact either, I think you’d find a lot of summaries that are off by a few words.