[Article] Technical rebuttal: semantic density and the “Naibbe” verbosity constraint

[magnesium]

I haven't got the letter yet, so this is based on the author's own summary on LinkedIn.

Taking the four devices in turn:

Abbreviation is the only one that clearly raises density. But it raises entropy too — abbreviation removes the redundant material that makes characters predictable from context. Voynichese sits well below plain Latin already. The one device that buys density moves h2 further from the target.

Permutation preserves h1 and should push h2 up toward it by destroying sequential regularity. A strictly deterministic permutation leaves structure intact but adds no concealment. Either way, permutation is length-preserving — it can't increase plaintext-per-glyph.

Positional behaviour puzzles me for a different reason. For position to carry semantic load you need the same glyph appearing contrastively in several slots, so that knowing the slot tells you something. Voynichese gives the opposite — knowing the glyph mostly tells you the slot.

Nulls lower entropy by adding predictable padding — but they lower density by definition, since they carry no plaintext. So the list contains one density-raising device (abbreviation) and one entropy-lowering device (nulls), and they work against each other.

This raises the key question: at what rate are nulls inserted, and is density computed per total glyph or per non-null glyph? If per non-null glyph, any density can be reached by declaring enough glyphs meaningless, while the entropy match is bought by the glyphs removed from the denominator.

Underneath all four is a point I'd expect the letter to address: no cipher can recover more information from a ciphertext than the ciphertext contains. The measured entropy of Voynichese is a ceiling on recoverable plaintext regardless of architecture. So was any text actually enciphered with the toy model, and are h1, h2, word-length distribution and type-token growth reported for the result?

Has anyone here got access to the full text yet?

Just to add to Torsten's point here: As I noted in the Naibbe cipher paper, the VMS's type length distribution and conditional character entropy jointly place tight constraints on how much plaintext information any given Voynichese token can feasibly contain, assuming that Voynichese glyphs or strings thereof are meant to map to individual plaintext letters. That means there is not a whole lot of budget for the full suite of operations Delgado proposes. No matter how you define a glyph within Voynichese, there are only so many Voynichese tokens that are eight or more glyphs long.

The permutation principle can in theory lower conditional entropy; the permutation could be alphabetizing every time, for example, and then you could encode which permutation restores the plaintext from the alphabetized letter string. But I have run a quick test of that on Naturalis Historia with a toy-model cipher (a two-letter code encoding which of 24 permutations is the correct decryption of an alphabetized four-letter chunk), and it doesn't lower entropy all the way to Voynich-like levels. Not only that, but it's just a bad cipher. It immediately becomes clear what the game is.