What if it's a "Lossy Encoder" for 15th-Century Latin?

[Pointless..]

my current working hypothesis on this idea is that that the glyph strings are not linguistic

Not linguistic or not phonetic?

If  Voynichese "words" mean things like leaf, flower, star or water then it's still a language. 
You may not read them loud as folium, flos, stella and aqua. You may not read them loud at all. But they remain words from some language.

Even if they mean things like "with water" or "from leaves" then they are still language. Such composite words are typical for agglutinative languages.

Of course it could be not existing language but constructed language. If I understand correctly that's what you are suggesting.

That's theoretically possible but I see several problems with it. If you develop your ideas further we could discuss it more.

Yes, of course, you are right that there is content in some linguistic form. That is, after all, the raison d'être of a book or manuscript.
What I am conjecturing is that the script may be mixed with some kind of markers or classifiers, distinct from ordinary diacritics, together with other additions that help guide the interpretation of the underlying linguistic content. So I wouldn't say it i a constructed language, I still think the actual content is in some European language.

Unfortunately, teaching commitments leave me with too little time at the moment to explore these ideas as deeply as I would like. I hope that will improve soon.

Below is a rough list of topics and possibilities that I have been thinking about. Please do not take them too literally or as fixed claims. Rather, consider them as meaningful tags pointing toward areas or concepts that may be worth exploring.

• The VMS is a lossy positional encoder for a bounded astro-medical domain — a structured record/schema, not a natural language or a cipher in the classic sense.
  
• Meaning is distributed across nested scopes: glyph → word → span → folio → section. Not completely decodable word-by-word.
  
• Glyphs are functionally typed (classifier / discriminator / structural marker / value), each class tied to a fixed positional slot.
  
• Capacity is multiplicative across these typed fields — so raw character entropy looks low but actual information density is higher than it appears.
  
• Immediate word repetition = fixed-width padding for underspecified schema entries, not meaningful redundancy.
  
• Currier A/B collapses to a per-folio boolean parameter modulating a shared template (leaning on Parisel).
  
• Autocopy is a null model / benchmark, not a competing theory — i.e. comparing autocopy's P(wi|wi₋₁) against the encoder's P(wi|s,c,d,t)
  

If anyone is aware of research, discussions, or even speculative work that touches on any of these topics, I would be very grateful for pointers.
As always, constructive comments, alternative perspectives, and critical observations are most welcome.