19-09-2025, 08:49 PM
Hi Tavi, thanks for the detailed feedback, I see where your skepticism is coming from.
You see a new member 0 previous posts or replies account created today and dropping a 31 page pdf. I would be sceptical as well.
Just to clarify: my work is not an LLM output. The framework doesn't rely on a model to invent content. Instead, I use EVA transcription as a baseline and then assign identifiers by a simple rule, based on the word's position relative to imagery.
For example:
A word that consistently appears near the stem of a plant is labeled a Stem-word.
A word consistently next to a jar-like image is labeled a Jar-word.
If a word modifies another label, it might be an Action-modifier like "Mixing-Action".
These identifiers are fully reproducible. Anyone starting with the EVA transcription and the folio images can follow these same rules to arrive at the same labels. The only role for AI is as a helper for checking consistency and spotting repetitions—not for generating the labels or their meanings.
As for the identical EVA words being mapped differently: that's a valid point, and it's a core feature of the method. The sampler uses context-based labeling. A word doesn't have a single, fixed meaning. It functions within its illustrated context. So the same EVA word can have different identifiers on different folios because its position and imagery have changed. This isn't a bug; it's a key part of how the framework operates.
On the Word "otolshade"
You are right to question the origin of words you haven't seen before. My documentation may have been unclear, but all the words listed are indeed EVA transcriptions found in the manuscript.
The word otolshade appears in my provided tables for both Folio 78v and Folio 85r. Its temporary identifier changes to reflect its visual context:
On Folio 78v (Balneological), it's identified as Depth-Shading because it's linked to the shading in the pools.
On Folio 85r (Cosmological), it's identified as Highlight-Zone because it's linked to shaded areas near diagram nodes.
These functional identifiers are consistent with my methodology and are derived directly from the visual evidence on each folio.
On Reproducibility
You stated that the methodology is unfollowable. The core of my method is based on visual and manual segmentation.
Visual Segmentation: I look at the glyph clusters and draw boundaries based on visual spacing.
EVA Transcription: I then apply a standard EVA transcription to each cluster.
Sequential Numbering: Each cluster is then assigned a sequential W number (W1, W2, etc.) for that specific folio.
This process is entirely manual and can be replicated by anyone. For example, on Folio 33v, the very first visual word cluster is designated F33v-W1 and transcribed as daiin, which is then given the temporary identifier Stem-Central due to its visual position. This process continues for every cluster, creating a transparent, traceable path from the manuscript image to the data table
I completely understand the skepticism given the number of "ChatGPT finds the secret code" posts out there. My goal here is to show a structured, repeatable way of analyzing function, not to push a machine-generated narrative.
If you'd like, I can create a step-by-step walkthrough showing how to reproduce the labels for a simple folio like 33v. Dm me and ill send it to you.
You see a new member 0 previous posts or replies account created today and dropping a 31 page pdf. I would be sceptical as well.
Just to clarify: my work is not an LLM output. The framework doesn't rely on a model to invent content. Instead, I use EVA transcription as a baseline and then assign identifiers by a simple rule, based on the word's position relative to imagery.
For example:
A word that consistently appears near the stem of a plant is labeled a Stem-word.
A word consistently next to a jar-like image is labeled a Jar-word.
If a word modifies another label, it might be an Action-modifier like "Mixing-Action".
These identifiers are fully reproducible. Anyone starting with the EVA transcription and the folio images can follow these same rules to arrive at the same labels. The only role for AI is as a helper for checking consistency and spotting repetitions—not for generating the labels or their meanings.
As for the identical EVA words being mapped differently: that's a valid point, and it's a core feature of the method. The sampler uses context-based labeling. A word doesn't have a single, fixed meaning. It functions within its illustrated context. So the same EVA word can have different identifiers on different folios because its position and imagery have changed. This isn't a bug; it's a key part of how the framework operates.
On the Word "otolshade"
You are right to question the origin of words you haven't seen before. My documentation may have been unclear, but all the words listed are indeed EVA transcriptions found in the manuscript.
The word otolshade appears in my provided tables for both Folio 78v and Folio 85r. Its temporary identifier changes to reflect its visual context:
On Folio 78v (Balneological), it's identified as Depth-Shading because it's linked to the shading in the pools.
On Folio 85r (Cosmological), it's identified as Highlight-Zone because it's linked to shaded areas near diagram nodes.
These functional identifiers are consistent with my methodology and are derived directly from the visual evidence on each folio.
On Reproducibility
You stated that the methodology is unfollowable. The core of my method is based on visual and manual segmentation.
Visual Segmentation: I look at the glyph clusters and draw boundaries based on visual spacing.
EVA Transcription: I then apply a standard EVA transcription to each cluster.
Sequential Numbering: Each cluster is then assigned a sequential W number (W1, W2, etc.) for that specific folio.
This process is entirely manual and can be replicated by anyone. For example, on Folio 33v, the very first visual word cluster is designated F33v-W1 and transcribed as daiin, which is then given the temporary identifier Stem-Central due to its visual position. This process continues for every cluster, creating a transparent, traceable path from the manuscript image to the data table
I completely understand the skepticism given the number of "ChatGPT finds the secret code" posts out there. My goal here is to show a structured, repeatable way of analyzing function, not to push a machine-generated narrative.
If you'd like, I can create a step-by-step walkthrough showing how to reproduce the labels for a simple folio like 33v. Dm me and ill send it to you.