Comment by brgsk

good catch - the example is sloppy. the real issue is lost-in-the-middle on long transcripts: the extracting model attends worse to material between endpoints, so "the transcript is still there" doesn't mean the extraction sees it equally.