(16-02-2026, 12:59 AM)nablator Wrote: You are not allowed to view links. Register or Login to view.If the cipher is as heavily homophonic as the one oshfdk suggests, the statistics are flattened and useless.

Frequencies of removed and added EVA character (max. 1, _ = 0) in paragraphs of RF1b:
0.1929 __
0.0389 d_
0.0379 _d
0.0365 e_
0.0316 _e
0.0278 q_
0.0260 rl
0.0214 _q
0.0212 kt
0.0206 sc
0.0200 lr
0.0198 tk
0.0173 oa
0.0173 i_
0.0171 _l
0.0163 cs
0.0161 _o
0.0157 o_
0.0151 l_
0.0121 y_
0.0121 t_
0.0113 _i
0.0111 ed
0.0111 _t
0.0107 s_
0.0105 _k
0.0103 _y
0.0099 ao
0.0087 oy
0.0083 rm
0.0083 k_
0.0071 kd
0.0071 dk
0.0065 ql
0.0062 de
0.0054 yo
0.0050 dt
0.0046 eq
0.0044 eo
0.0042 r_
0.0042 ol
0.0042 lm
0.0042 _r
0.0038 qe
0.0038 oe
0.0038 el
0.0036 sy
0.0036 ld
...

(16-02-2026, 04:30 AM)Dunsel Wrote: You are not allowed to view links. Register or Login to view.Michael Gresko published a paper late last year on the only cypher I've see that comes even marginally close to the Voynich. I know he has a post on here but couldn't find it.

You are not allowed to view links. Register or Login to view.

That's a link to his paper

You are not allowed to view links. Register or Login to view.

And that's to his github where you can download his python files. 

May not be what you're looking for but it may give you a few hints.

(16-02-2026, 07:27 PM)Fontanellean Wrote: You are not allowed to view links. Register or Login to view.Yes, I like the Naibbe cipher. I would like to see what happens to the long-range correlations if an actual card-drawing process is implemented, because Greshko's implementation from which he drew the statistics in his paper actually treats the table selections as independent events. He seems to think it wouldn't make much difference to use an actual card-draw, and maybe he's right, but I would still like to see the degree of the effect.

(16-02-2026, 08:38 PM)magnesium Wrote: You are not allowed to view links. Register or Login to view.

(16-02-2026, 07:27 PM)Fontanellean Wrote: You are not allowed to view links. Register or Login to view.Yes, I like the Naibbe cipher. I would like to see what happens to the long-range correlations if an actual card-drawing process is implemented, because Greshko's implementation from which he drew the statistics in his paper actually treats the table selections as independent events. He seems to think it wouldn't make much difference to use an actual card-draw, and maybe he's right, but I would still like to see the degree of the effect.

I agree that it would be interesting to test this further. To make long-range correlations work in the context of a homophonic substitution cipher without assuming a fine-tuned plaintext structure (i.e., let's assume a plaintext Hurst exponent of 0.5), I suspect that there need to be long-range correlations in the table selection process that are reliably carried by the cipher's homophone inventories. That is to say, each "table" consists of structurally distinctive glyph strings, so if there are long-range correlations in the table selection sequence, the resulting glyph strings reliably bake those correlations into the ciphertext.

Reconciling this with the current Naibbe cipher probably would require revising the tables, at a minimum. The bigram prefix and suffix inventories within the current Naibbe cipher's tables were populated specifically assuming random respacing and letter-by-letter table selection. So even if long-range correlations were imposed on the table selection sequence within the current cipher, I'm not sure that the resulting ciphertext would have strong long-range correlations, as the current inventories of bigram prefixes and suffixes would essentially act as random noise.

(16-02-2026, 09:20 PM)Fontanellean Wrote: You are not allowed to view links. Register or Login to view.

(16-02-2026, 08:38 PM)magnesium Wrote: You are not allowed to view links. Register or Login to view.

(16-02-2026, 07:27 PM)Fontanellean Wrote: You are not allowed to view links. Register or Login to view.Yes, I like the Naibbe cipher. I would like to see what happens to the long-range correlations if an actual card-drawing process is implemented, because Greshko's implementation from which he drew the statistics in his paper actually treats the table selections as independent events. He seems to think it wouldn't make much difference to use an actual card-draw, and maybe he's right, but I would still like to see the degree of the effect.

I agree that it would be interesting to test this further. To make long-range correlations work in the context of a homophonic substitution cipher without assuming a fine-tuned plaintext structure (i.e., let's assume a plaintext Hurst exponent of 0.5), I suspect that there need to be long-range correlations in the table selection process that are reliably carried by the cipher's homophone inventories. That is to say, each "table" consists of structurally distinctive glyph strings, so if there are long-range correlations in the table selection sequence, the resulting glyph strings reliably bake those correlations into the ciphertext.

Reconciling this with the current Naibbe cipher probably would require revising the tables, at a minimum. The bigram prefix and suffix inventories within the current Naibbe cipher's tables were populated specifically assuming random respacing and letter-by-letter table selection. So even if long-range correlations were imposed on the table selection sequence within the current cipher, I'm not sure that the resulting ciphertext would have strong long-range correlations, as the current inventories of bigram prefixes and suffixes would essentially act as random noise.

I have a lot to learn about the mathematics of this, but I don't understand why that should be the case. So long as the underlying letters that your tables are representing have unequal distribution, which in the plaintext they do, then shouldn't the correlation be able to show through even if the glyphs have been randomly assigned? Or is it just that it's far too weak?