Notes from the work · 2026
Blog
Engineering notes, dataset releases, and method write-ups—including the runs that failed and the decisions that changed afterward.
Structuring 794 Sanskrit books
An agent produced 4,054,517 correctly linked records and still scored zero. The failure was in the delivery contract.
The verifier is the product
Why structural confidence comes from the checks a benchmark enforces, not the fluency of the artifact it produces.
601,780 verses with stable IDs
The first dataset release built around addresses that survive normalization, rebuilding, and new editions.
09 Jul 2026Sandhi splitting without a parserSanskrit & Method
27 Jun 2026Parsing GRETIL’s 794 pagesEngineering
11 Jun 2026What a Sanskrit search index has to knowEngineering
30 May 2026Twenty dictionaries, one sense graphDataset release
14 May 2026Why we do not auto-translateSanskrit & Method
16 Apr 2026A passage ID scheme we will not regretEngineering