@ nablator
Okay, now I understand what you mean.
Yes, I've already done that kind of calculation. For each of the 7 line-start markers as I define them (others work with fewer): o, qo, d, y, s, p, t - and ch / sh, I tested many properties of the remaining text in the line: gallow density, various endings, token length, line length, daiin rate, etc.
I deliberately limited this to text folios and omitted the diagram pages (f57v, f67–74 Zodiac, f85–86, fRos). At first, I had some incredibly good results, but these turned out to be artifacts from labels and other "contaminants" from those pages (sometimes you can still see this effect showing through in other calculations here in this forum).
[
attachment=15234]
The numbers are cross-checked, but I welcome any further verification. Who or what is ever error-free
.
p-start is by far the most striking: 10 "global" properties shifted at once. p-start lines are longer, have longer tokens, significantly more gallows, more y-endings, less daiin, fewer bank-gallows, fewer repetitions.
But how to interpret this? The statistical background needs more careful thought than I've given it so far - without your question I wouldn't have published this table yet. For example, lines that have lots of gallows in them anyway might start more often with a "p" gallow, etc. So one has to be cautious with the interpretation - because "t" shows the same gallow effect as p, only weaker - which is a bit suspicious, or maybe a hint at some kind of "functionality" - who knows
qo and o go in the other direction - shorter lines or shorter tokens in the rest. ch and s also have clear signatures.
I find the ch signals particularly interesting, I've done some additional tests on those but that would go too far here.
(I'm sure everyone has known this all along, but for some reason it only recently fully clicked for me that "ch" is a glyph variant - and what that implies. The bench gallows are really just "ch" variants... I'd been thinking of them as gallow variants...)
What the table really means - cipher key, different text types, some other process - I don't know yet. But the rules exist.
That said, one should handle this data with some caution...
PS: I mentioned the bigrams: there are some notable patterns, but they are even more heavily skewed by statistical side-effects, so I don't want to publish them - they#d give a misleading impression. For example, if tokens following "p" are longer on average, then naturally more bigrams will appear after "p" as well, regardless of any real structural link.
The only genuinely noteworthy patterns I could isolate are that an "s" beginning attracts "ed," while a "qo" beginning repels "ar."
