05-05-2026, 10:07 AM
So here is the proposed multilayer mechanism (more details to come in my github repo):
Each word containing ch or sh followed immediately by o or e is mapped to a template by replacing the post-digraph vowel with a placeholder X.
For example, chody and chedy both map to chXdy; shol and shel both map to shXl.
Words containing multiple substitution sites have each site replaced independently.
Each word token contributes one count to its template in the appropriate folio-state column, and is scored as cho or che at each substitution site.
The cho rate for a template in a given state is the number of cho substitution events divided by the total substitution events for that template in that state.
We retain only templates with at least 10 substitution events in both folio states. Of the hundreds of distinct templates in the corpus, 31 meet this criterion.
These 31 templates account for 4,085 substitution events, comprising the large majority of all classifiable tokens.
Here are the templates:
+----------------+--------------+---------+------+---------+------+---------+
| Class | Template | Rate₁ | n₁ | Rate₀ | n₀ | |Δ| |
+----------------+--------------+---------+------+---------+------+---------+
| Fixed cho | chXl | 1.000 | 217 | 0.948 | 135 | 0.052 |
| Fixed cho | shXl | 1.000 | 102 | 0.953 | 64 | 0.047 |
| Fixed cho | shXr | 0.966 | 58 | 0.935 | 31 | 0.030 |
+----------------+--------------+---------+------+---------+------+---------+
| Fixed che | shXy | 0.098 | 41 | 0.033 | 304 | 0.065 |
| Fixed che | chXey | 0.000 | 26 | 0.000 | 156 | 0.000 |
| Fixed che | chXol | 0.031 | 32 | 0.008 | 123 | 0.023 |
| Fixed che | shXey | 0.000 | 27 | 0.000 | 108 | 0.000 |
| Fixed che | shXol | 0.000 | 23 | 0.000 | 72 | 0.000 |
| Fixed che | chXor | 0.075 | 40 | 0.000 | 47 | 0.075 |
| Fixed che | chXody | 0.000 | 11 | 0.000 | 63 | 0.000 |
| Fixed che | chXo | 0.067 | 15 | 0.000 | 46 | 0.067 |
| Fixed che | shXo | 0.000 | 15 | 0.034 | 29 | 0.034 |
| Fixed che | otchXy | 0.000 | 12 | 0.032 | 31 | 0.032 |
| Fixed che | shXor | 0.000 | 13 | 0.000 | 28 | 0.000 |
| Fixed che | okchXy | 0.067 | 15 | 0.000 | 22 | 0.067 |
+----------------+--------------+---------+------+---------+------+---------+
| Switchable | chXdy | 0.921 | 38 | 0.116 | 362 | 0.805 |
| Switchable | shXdy | 0.870 | 23 | 0.096 | 271 | 0.774 |
| Switchable | shX | 0.880 | 92 | 0.458 | 48 | 0.422 |
| Switchable | chXky | 0.703 | 37 | 0.210 | 62 | 0.493 |
| Switchable | chXs | 0.778 | 18 | 0.387 | 62 | 0.391 |
| Switchable | chX | 0.976 | 41 | 0.529 | 34 | 0.446 |
| Switchable | chXty | 0.812 | 32 | 0.483 | 29 | 0.330 |
| Switchable | chXdaiin | 1.000 | 21 | 0.436 | 39 | 0.564 |
| Switchable | chXcthy | 0.917 | 12 | 0.167 | 36 | 0.750 |
| Switchable | shXky | 0.500 | 10 | 0.111 | 27 | 0.389 |
| Switchable | shXdaiin | 0.917 | 12 | 0.500 | 16 | 0.417 |
+----------------+--------------+---------+------+---------+------+---------+
| Intermediate | chXy | 0.213 | 89 | 0.021 | 426 | 0.192 |
| Intermediate | chXr | 0.962 | 158 | 0.898 | 49 | 0.064 |
| Intermediate | chXar | 0.158 | 19 | 0.043 | 46 | 0.114 |
| Intermediate | chXal | 0.182 | 11 | 0.000 | 31 | 0.182 |
| Intermediate | kchXy | 0.143 | 14 | 0.000 | 14 | 0.143 |
+----------------+--------------+---------+------+---------+------+---------+
(Rate1 and Rate0 give the cho rate in O and E-folios. |∆| is the absolute rate difference between states.).
Templates are classified into four categories based on their cho rate in each folio state:
Fixed cho (F1): rate > 0.9 in both states. 3 templates, 607 events.
Fixed che (F0): rate < 0.1 in both states. 12 templates, 1,299 events.
Switchable (S): absolute difference between states ≥ 0.2. 11 templates, 1,322 events.
Intermediate (I): does not meet any of the above criteria. 5 templates, 857 events.
How does it work?
Imagine each folio of the manuscript has a switch X that can be set to one of two positions: O or E.
Now suppose we look at a folio where the switch is set to E. What words will we find?
chol and chel? Mostly chol, just like in all folios. Their template chXl is fixed cho (Rate = 1.000 in O-folios, 0.948 in E-folios).
The template structure drives the vowel choice here: the switch has almost no effect.
shody and shedy? Mostly shedy. Their template shXdy is switchable (Rate = 0.870 in O-folios, 0.096 in E-folios).
In E-folios, shody is rare; in O-folios, shody dominates. This is where the switch has its strongest effect.
choey and cheey? Only cheey, never choey, regardless of the switch position.
Their template chXey is fixed che with a rate of exactly 0.000 in both states. This is an absolute rule that the switch cannot override.
chool and cheol? Only cheol, never chool, again regardless of the switch.
Their template chXol is fixed che (Rate = 0.031 in O-folios, 0.008 in E-folios).
The system appears to avoid placing two o vowels adjacent across the digraph boundary.
The switch does not operate uniformly: it strongly modulates some word contexts (the switchable templates) while leaving others
completely untouched (the fixed templates). The template (the consonantal environment surrounding the substitution site) determines
whether the switch has any power at all.
This aligns nicely with Jorge's intuition that the A/B split could be two encodings of the same underlying mechanism.
But the full picture is more nuanced. There are in fact two independent signals hiding inside the Currier A/B distinction:
A strong bimodal signal (the one we just explained), based on the cho/che alternation. This is a discrete boolean switch set once per folio.
A much weaker gradient signal carried primarily by the d/l pair (and possibly others, which I have not yet attempted to identify).
This is not discrete: 70.6% of folios fall in the middle range, and the bimodality coefficient (0.319) is well below the threshold for bimodality. It behaves like a continuous dial, not an on/off switch.
Currier's A/B classification conflates these two signals into a single binary label, which is why it works reasonably well but leaves 71% of inter-folio variance unexplained.
To me, the most intriguing finding is that these two signals are completely orthogonal. Within the Herbal section, the correlation between the cho/che ratio and the d/l ratio is essentially zero. Knowing the switch state of a folio tells you nothing about its d/l ratio, and vice versa. They are independent dimensions of the manuscript's structure, driven by separate mechanisms that happen to partially align in Currier's coarse classification.
Each word containing ch or sh followed immediately by o or e is mapped to a template by replacing the post-digraph vowel with a placeholder X.
For example, chody and chedy both map to chXdy; shol and shel both map to shXl.
Words containing multiple substitution sites have each site replaced independently.
Each word token contributes one count to its template in the appropriate folio-state column, and is scored as cho or che at each substitution site.
The cho rate for a template in a given state is the number of cho substitution events divided by the total substitution events for that template in that state.
We retain only templates with at least 10 substitution events in both folio states. Of the hundreds of distinct templates in the corpus, 31 meet this criterion.
These 31 templates account for 4,085 substitution events, comprising the large majority of all classifiable tokens.
Here are the templates:
+----------------+--------------+---------+------+---------+------+---------+
| Class | Template | Rate₁ | n₁ | Rate₀ | n₀ | |Δ| |
+----------------+--------------+---------+------+---------+------+---------+
| Fixed cho | chXl | 1.000 | 217 | 0.948 | 135 | 0.052 |
| Fixed cho | shXl | 1.000 | 102 | 0.953 | 64 | 0.047 |
| Fixed cho | shXr | 0.966 | 58 | 0.935 | 31 | 0.030 |
+----------------+--------------+---------+------+---------+------+---------+
| Fixed che | shXy | 0.098 | 41 | 0.033 | 304 | 0.065 |
| Fixed che | chXey | 0.000 | 26 | 0.000 | 156 | 0.000 |
| Fixed che | chXol | 0.031 | 32 | 0.008 | 123 | 0.023 |
| Fixed che | shXey | 0.000 | 27 | 0.000 | 108 | 0.000 |
| Fixed che | shXol | 0.000 | 23 | 0.000 | 72 | 0.000 |
| Fixed che | chXor | 0.075 | 40 | 0.000 | 47 | 0.075 |
| Fixed che | chXody | 0.000 | 11 | 0.000 | 63 | 0.000 |
| Fixed che | chXo | 0.067 | 15 | 0.000 | 46 | 0.067 |
| Fixed che | shXo | 0.000 | 15 | 0.034 | 29 | 0.034 |
| Fixed che | otchXy | 0.000 | 12 | 0.032 | 31 | 0.032 |
| Fixed che | shXor | 0.000 | 13 | 0.000 | 28 | 0.000 |
| Fixed che | okchXy | 0.067 | 15 | 0.000 | 22 | 0.067 |
+----------------+--------------+---------+------+---------+------+---------+
| Switchable | chXdy | 0.921 | 38 | 0.116 | 362 | 0.805 |
| Switchable | shXdy | 0.870 | 23 | 0.096 | 271 | 0.774 |
| Switchable | shX | 0.880 | 92 | 0.458 | 48 | 0.422 |
| Switchable | chXky | 0.703 | 37 | 0.210 | 62 | 0.493 |
| Switchable | chXs | 0.778 | 18 | 0.387 | 62 | 0.391 |
| Switchable | chX | 0.976 | 41 | 0.529 | 34 | 0.446 |
| Switchable | chXty | 0.812 | 32 | 0.483 | 29 | 0.330 |
| Switchable | chXdaiin | 1.000 | 21 | 0.436 | 39 | 0.564 |
| Switchable | chXcthy | 0.917 | 12 | 0.167 | 36 | 0.750 |
| Switchable | shXky | 0.500 | 10 | 0.111 | 27 | 0.389 |
| Switchable | shXdaiin | 0.917 | 12 | 0.500 | 16 | 0.417 |
+----------------+--------------+---------+------+---------+------+---------+
| Intermediate | chXy | 0.213 | 89 | 0.021 | 426 | 0.192 |
| Intermediate | chXr | 0.962 | 158 | 0.898 | 49 | 0.064 |
| Intermediate | chXar | 0.158 | 19 | 0.043 | 46 | 0.114 |
| Intermediate | chXal | 0.182 | 11 | 0.000 | 31 | 0.182 |
| Intermediate | kchXy | 0.143 | 14 | 0.000 | 14 | 0.143 |
+----------------+--------------+---------+------+---------+------+---------+
(Rate1 and Rate0 give the cho rate in O and E-folios. |∆| is the absolute rate difference between states.).
Templates are classified into four categories based on their cho rate in each folio state:
Fixed cho (F1): rate > 0.9 in both states. 3 templates, 607 events.
Fixed che (F0): rate < 0.1 in both states. 12 templates, 1,299 events.
Switchable (S): absolute difference between states ≥ 0.2. 11 templates, 1,322 events.
Intermediate (I): does not meet any of the above criteria. 5 templates, 857 events.
How does it work?
Imagine each folio of the manuscript has a switch X that can be set to one of two positions: O or E.
Now suppose we look at a folio where the switch is set to E. What words will we find?
chol and chel? Mostly chol, just like in all folios. Their template chXl is fixed cho (Rate = 1.000 in O-folios, 0.948 in E-folios).
The template structure drives the vowel choice here: the switch has almost no effect.
shody and shedy? Mostly shedy. Their template shXdy is switchable (Rate = 0.870 in O-folios, 0.096 in E-folios).
In E-folios, shody is rare; in O-folios, shody dominates. This is where the switch has its strongest effect.
choey and cheey? Only cheey, never choey, regardless of the switch position.
Their template chXey is fixed che with a rate of exactly 0.000 in both states. This is an absolute rule that the switch cannot override.
chool and cheol? Only cheol, never chool, again regardless of the switch.
Their template chXol is fixed che (Rate = 0.031 in O-folios, 0.008 in E-folios).
The system appears to avoid placing two o vowels adjacent across the digraph boundary.
The switch does not operate uniformly: it strongly modulates some word contexts (the switchable templates) while leaving others
completely untouched (the fixed templates). The template (the consonantal environment surrounding the substitution site) determines
whether the switch has any power at all.
This aligns nicely with Jorge's intuition that the A/B split could be two encodings of the same underlying mechanism.
But the full picture is more nuanced. There are in fact two independent signals hiding inside the Currier A/B distinction:
A strong bimodal signal (the one we just explained), based on the cho/che alternation. This is a discrete boolean switch set once per folio.
A much weaker gradient signal carried primarily by the d/l pair (and possibly others, which I have not yet attempted to identify).
This is not discrete: 70.6% of folios fall in the middle range, and the bimodality coefficient (0.319) is well below the threshold for bimodality. It behaves like a continuous dial, not an on/off switch.
Currier's A/B classification conflates these two signals into a single binary label, which is why it works reasonably well but leaves 71% of inter-folio variance unexplained.
To me, the most intriguing finding is that these two signals are completely orthogonal. Within the Herbal section, the correlation between the cho/che ratio and the d/l ratio is essentially zero. Knowing the switch state of a folio tells you nothing about its d/l ratio, and vice versa. They are independent dimensions of the manuscript's structure, driven by separate mechanisms that happen to partially align in Currier's coarse classification.