VMS to pseudo - latin ( not a solution !)

[MarcoP]

Thanks, Marco, that's quite illuminating. I'd still say that this isn't too far from what solvers tend to do.

I agree this is very close to the pre-LLMs solutions that were popular until recently.

I think there are ways the results could be improved eg:

• Use a transliteration system like CUVA that aggregates the most common EVA bigrams (this marginally increases entropy).
  
• More sophisticated simple substitution, e.g. instead of simply matching character frequencies, optimize bigram frequencies (I guess this would mostly map EVA:o,a,e into Vowels).
  
• Use the frequency of target-language words together with Levenshtein distance, so that most frequent words are preferred.
  

In any case, solvers tend not to be very sophisticated, and it's possible that these "improvements" would make the system less close to the solutions we are familiar with.