12-11-2026, 10:28 PM
Mauro wrote:But I guess compiling the tables bifolio by bifolio could improve the scores? Maybe not (less words translates to more unreliable statistics) and surely it could be regarded as overfitting, but why not give it a try?
Good idea, I tryed it. It doesnt help.
I built the skribe's glyph tables from each bifolio alone, plus two checks: each page written with the table of the other pages of its bifolio (so it never sees the page it writes), and a control where each bifolio uses the table of another bifolio of the same hand and section.
| tables | tests passed (of 89) | same bifolio vocabulary (real 1.30) |
| hand x section (#81) | 80.8 | 1.48 |
| per bifolio | 78.2 | 1.95 |
| per bifolio, leave one page out | 82.0 | 1.88 |
| another bifolio (control) | 76.4 | 1.97 |
(counted per text, average of 5, thats why it is lower than in #81)
The bifolio tables make the four pages too similar to each other, and the control does the same with a different bifolio. So the bifolio is a light push on top of the hand, not its own glyph habits, which is what the random "mood" per bifolio already does.


![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)