So how can you verify correctness of transcription and summary in a way that is repeatable over time?

Agreed. That sounds like a recipe for "we don't know how 'the algorithm' came up with what it did" kinds of excuses when, inevitably, inaccuracies are found. It also seems, conveniently, to make the processing system practically unimpeachable.