Experiments with Claude

There are 186 adhyāyas in the Suśrutasaṃhitā. Our project members have critically edited and translated about 23 of them. Clearly there is a lot still to do!

The chapters that have not been edited have been published online as collations. That means, we took one of the manuscript transcripts, often H, and cleaned it up in a very simple and generic manner, removing references to the actual manuscript text (deletions, insertions, errors). We called the resulting file a provisional edition. Saktumiva allows us to collate this provisional edition against the manuscript witnesses. Of course, the provisional, un-critically-edited edition will mostly agree with the witness (probably H) from which it was derived. Nevertheless, there is great value in this provisional edition. Even in such a crude form, it provides an excellent idea of what the Nepalese version of the Suśrutasaṃhitā says, and the collation information in the apparatus gives the reader the opportunity to think creatively about unresolved difficulties in the provisional edition.

I have been thinking recently about how to improve these provisional edition chapters. At the same time, I have been learning about the emerging capabilities of AI systems like Gemini, ChatGPT and Claude.

I decided to experiment with the following idea. Ask Claude to study the 23 critically edited chapters alongside the witnesses that underly them. Derive a set of rules or patterns that capture our editorial practices. Then read the remaining un-critically-edited chapters and apply those rules to them.

The preliminary results of Claude’s output for Suśrutasaṃhitā, Sūtrasthāna 6 (Ṛtucaryā) can be seen here:

• https://sushrutaproject1.github.io/Sushrutasamhita_Saktumiva2/01-su.su-1-31/provisional-edition_sutrasthana-06.xml

You can compare this with the “raw” collation in the online text of Sūtrasthāna 1-31:

• https://sushrutaproject1.github.io/Sushrutasamhita_Saktumiva2/01-su.su-1-31/provisional-edition_sutrasthana-1-31.xml

This is just the start of exploring whether AI can do serious critical editing work that will help with Sanskrit manuscript traditions where there are very large texts, very high numbers of manuscripts, or both.