The problem is also a sample bias I think. He clearly hand picked low probability words as evidence. In other words, the probability that there exists a subset

The problem is also a sample bias I think. He clearly hand picked low probability words as evidence. In other words, the probability that there exists a subset of words in a pair of papers that are "globally rare" among cryptography papers is likely high. So to answer P(Nick|Evidence) we can't use ~P(selected subset|crypto), we better use ~P(all words|crypto) and compare that to all other plausible sources. As a clarifying anecdote, remember low probability events happen all the time if the sample space is simply large. Throw a dice ~8 times in a row an the probability of that event is lower than winning the lottery or getting hit by lightning, yet it happens every time. If you pick every word from the source paper and multiply their relative frequencies out, you will get a number astronomically low, and that happens for any paper.