Evidence for 2 rules linking words in Herbal A->B

[npcompl33t]

I'm seeing quite a lot of suggestive evidence that Herbal A's -ol and -or material is preserved in Herbal B through some type of splitting, compounding, or merging.

For example, the words containing *aiin interact with chol and chor, and -or/-ol, in unexpected ways.

• In herbal A, all cho forms are more likely to appear before an aiin form and less likely to appear after it. Chol is the most likely word to proceed *aiin.
  

• In herbal B, or specifically occurs at 2.4x the base rate in front of *aiin. All other possible variants, ol, or, ar, and even 4 of the 6 chors are more likely to appear AFTER an *aiin word.
  

This seems to suggest that some material is actually being reorganized across words, not merely transformed. For example, this material could be getting reorganized in B in the following ways: 

• chol daiin → olaiin
  
• chor daiin → or aiin 
  
• chor daiin → aiin chor
  
• chol daiin → chdy olaiin
  
• chor daiin → chedaiin or
  
• okol daiin → okaiin dol
  

My working theory to explain the missing -ol and -or mass is that not every word can be 'translated' into a different section with a 1:1 transformation. For at least some contexts, the content from multiple words in the source is reorganized across multiple words in the destination.

I've already been able to improve my existing rule-set by extending it to consider preceding and following words. 

• With respect to determining whether a word should begin with da- (like daiin) or a- (like aiin, air, etc), the preceding ending strongly conditions its form. r+s endings on the previous word make the a- 22x more likely, l makes it 0.5x as likely, y makes it 0.15x as likely. 
  
• this means that technically the following word determines whether -al or -edy appears. an a initial (like aiin) makes the -al ending on the prior word more likely, d- (like daiin) makes the -edy form more likely.
  

There are also indicators that words beyond the immediate adjacent two influence form realization, and before anyone yells 'self-citation', these gains cannot all be explained by an exact or similar copy of a word on the same line. 

In Herbal B, edy is more likely than -al to appear on lines whose other, nonadjacent words have fewer o-initial and l/r-final forms, and more y-final forms. I see predictive gains around .076 bits/token, which persists across several controls. This does not seem to affect all spelling choices, for instance bare gallows. o-, and qo- fronts only see gains of .006 bits/token. These results are suggestive but the actual mechanism is still unresolved.

I'm also seeing indicators that other words possessing an initial q can influence whether another word on the same line takes an -edy form. The actual position within the line of this other q- word matters. Words containing q- to the left of the target word, seems to make -edy forms more likely to occur, and this survives withholding the target word stem, the physical sheet, and removing repeated stem family words.

A concrete example of this can be seen in the Herbal B line:

• qokar okar [okedy] dar ykchey kaiin okas chedy okar a ralos
  

Moving the sole q from the first qokar to the later identical okar preserves every core, both immediate neighbors and total q- words, yet the held-out model’s edy probability falls .805 → .544.

These are mostly exploratory and suggestive findings. Held out models can boil down information in complex ways, and this definitely needs more in depth analysis and testing to validate and pin down what is happening mechanically. 

I will unfortunately not be able to work on this much over the coming week, but very excited about some of these findings and their potential to explain more correspondence between A/B.