if you crib for many words you can decode the message you are hoping for. That seems like the bias in machine learning. So you try to crib for words until the message seems appropriate for you, but maybe the decryption is not correct. People have an extraordinary capacity to connect words to make it sound like there is some real meaning. But the decrypted words were not naturally found by the algorithm but suggested by the authors. To test my hypothesis, about the lack of credibility of such a method of decrypting, I suggest to decode the same message thinking that the author of the plaintext is someone who got rich by a lottery prize. You may try millions crib like "I like millions", "I am rich" "I will win the lottery", and the like. And finally you will communicate to your audience only the few words that work for you to "decrypt" the message. I think the result will be something along the lines of: You finally won the big prize, the millions are yours, congratulations.

What I am saying is trivially true for a very short plaintext, think of txyz and invent any work you like. So there must be a trade-off between the space of solutions and the message length, a signal to noise ratio. The solution propose 3 part of the z340, and each one add new capacity to the channel of solutions. Perhaps something like the VC dimension. caveat: I am not a cryptographer.

Perhaps something like: https://www.mdpi.com/1099-4300/22/4/438/htm