Tested K&A at scale, found something, need your help

[CorwinFr]

Are you still using AI LLMs?

Hello tavie,

Yes I use Claude Code for the computational work, I'm not really coding myself anymore.

This thread is about my previous paper wich was wrong. The phonetic parser was returning valid Latin words almost every time, giving a false positive feedback loop...

The new work is different, no parser, no phonetic assumption. It's purely structural, more like studying how a pharmacist writes shorthand, the way they write ASA for acetylsalicylic acid or APAP for acetaminophen. Arbitrary codes memorized through daily practice.

From this idea I see each herbal page as possibly one entry in the pharmacist's personal codebook. The first word on the page could be his shorthand for that plant, like ASA is our code for aspirin. Folio You are not allowed to view links. Register or Login to view. starts with pched, You are not allowed to view links. Register or Login to view. with pcheod, You are not allowed to view links. Register or Login to view. with foch, etc.

What's interesting is that these roots seem to reappear embedded inside compound words in the recipe section: o.pched.y could mean something like "with [f26v plant] in...". Same root, different context. As if the herbal section is his reference dictionary and the pharma section uses it.

Now it's not perfect, 92% of herbal folios have a distinct first word but 5 words are shared by 2 or 3 folios. One case makes pharmacological sense tho, Sonchus and Hieracium (both Asteraceae, latex composites) share the same code tshor, wich is what you'd expect if a pharmacist considers them interchangeable. Two other pairs are on adjacent folios. The last two I can't explain.

53% of pharma block-initial words (first word of each starred paragraph, 154 out of 286 blocks across f103r-f116v) contain a herbal root as substring. I might be wrong about what it means but the pattern is in the text, anyone can check it.

Under the hood I parsed the full ZL v3b transcription into a structured dataset where every word is decomposed into its components (root, prefix, suffix, logogram status). This lets me link any hypothesis to the actual text, test alternatives like K&A mappings or other transliteration schemes, and run statistical checks against the 8 medieval corpora I assembled. It's all on GitHub if anyone wants to play with it or break it.