What are the characteristics of Labelese?

[ReneZ]

I ran a simple test on 1000 Latin words:

• vowels were removed (with the exception of word-initial occurrences)
  
• the remaining characters were sorted alphabetically
  

So "manibus" is encoded as "bmns".

This results in 14% "collisions" (different source words being mapped into identical coded words), a number comparable with the overlaps in Voynichese single-word labels. Of course, this method also results in a considerable decrease in entropy values.

This is certainly true, but the character set is reduced, so not only are all entropy values reduced, also the 'theoretical maximum' entropy values are reduced.
It means that it is very difficult to compare entropy values for different sizes of character sets.

In the very early days of the internet I had done something very similar, in order to try to generate word patterns. I can't remember exactly, but I believe that I just rearranged all characters in every word to put them in alphabetical order. Maybe I also enforced strict vowel-consonant alternation (to the extent possible). The number of collisions was surprisingly small.
Some reduction in entropy was expected (the combinations become more limited) but also that reduction was surprisingly limited.