Can LAAFU effects be modeled?

[Mauro]

We'd need a well-balanced mix of things that reduces entropy and also breaks up standard phrases.

A possibility which came to my mind recently is a many-to-many substitution cipher. I don't like it at all, first because it's a practically unfalsifiable theory, and second because it makes the VMS undecipherable by construction (even if it would not be gibberish, after all). However I'd expect it to reduce n-grams and word entropies (say, 'et' 'aut' 'ab' could all be ciphered as 'aiin') and to break up standard phrases (because the same source word could be ciphered in many different ways). Just a fringe idea.