Filings reader

The bet

Analysts at small funds still read long documents by hand: 10-Ks, credit agreements, indentures, prospectuses. Large language models can now extract verbatim, cited answers from a 600-page filing for roughly a hundredth of what it costs to read it in context, and the extraction can be verified against the source mechanically.

The tool is the verifying reader: ask a question, get quotes with locations, have every quote re-checked against the document before it reaches a spreadsheet. Anything that becomes a number stays in deterministic code.

The gate

Before writing more code I set out to ask ten people who do this work one question:

What did you read by hand last week, and what did it cost you, in hours or in things you missed?

If fewer than three of ten named a document type and a cost, the idea would be dropped.

Where it comes from

A year of building a distressed-credit screener, where the bottleneck was never the math. It was that credit agreements have no structured tags and someone has to read them.

Where it stands

In October I moved on to put my time back into NodFirst.