Writing

A progress update: moving to bond indentures, reading 808 of them with a model on my laptop, a mistake I made about the ones it got wrong, a small classifier experiment that didn't go the way I expected, and the first scores against labels I've checked myself.

The covenant extractor: how it reads a credit agreement, how it's scored against labels I haven't reviewed yet, what 184 documents cost to run, and where it's still wrong.

Building a complete financial dataset out of primary sources: point-in-time fundamentals and full statements for every US filer, Treasury and Fed credit history, and a corpus of credit agreements underneath structured covenant terms that cite the exact line they came from.