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

[Pointless..]

Some observations:

You don't say anything about encoding (the used cipher which hypothetically turns let's  say "radix" into "chedy"), just about the structure of the source text.
Do you have any ideas about the cipher?

Do you think that if the text lacked basic words like "and", "so", "with" etc. but was encrypted in some simple way, then it would be harder to crack?
Personally I don't think so.

The structure of Voynichese isn't also compatible with what you suggest. It has a high percentage of hapax legomena (unique) words which is similar to Latin with declension and much higher than let's say English.

It has very common words like "daiin" which would be good candidates for words like "and" (but it doesn't work). If there is no "and" then what "daiin" could be?

Thank you for these comments, appreciated.
You are right, the Voynich manuscript appears to have an unusually hapax-rich vocabulary. Depending on the transcription and tokenization method, I understand that around 70% of its distinct space-delimited forms occur only once. This is quite high compared with ordinary prose of similar length, although the result depends heavily on whether Voynich spaces actually mark word boundaries and whether its glyph strings correspond to linguistic words at all, as my current working hypothesis on this idea is that that the glyph strings are not linguistic.

Under a lossy positional model, this anomalous hapax rate would sense. If Voynich spaces do not separate normal spoken words, but instead separate clusters of positional data or semantic parameters, the number of unique combinations will naturally explode.

Think of it like combining our earlier stripped Latin roots with specific disambiguation cues or positional padding. If a scribe combines a core root with a highly specific modifier to create a single space-delimited string, that exact string might only ever be needed once in the entire book. The 70% hapax rate isn't a sign of an impossibly large vocabulary; it is the statistical footprint of a modular, combinatorial system where tiny tweaks to a string create thousands of single-use variants.

And regarding the mysterious "daiin daiin", I think will test case for this whole writing system idea - it's the single token-pair where the padding hypothesis, the classifier typology, and the autocopy-null all make competing, checkable predictions.

How does this yet a bit vague fidea sound?