If I recall correctly, the client pre-computes hashes of random fragments of each file then "quizzes" the hosts on that information periodically. If they're not

If I recall correctly, the client pre-computes hashes of random fragments of each file then "quizzes" the hosts on that information periodically. If they're not actually storing a copy of the data, they won't be able to compute the resulting hash. (I may be misremembering some of the details, but I believe that's how it works in principle.)