For financial analysts
One document is chat.
The whole corpus is a tournament.
Claude can answer one document. Foundry takes that same question, runs it across all the documents, and turns the answers into a dataset with proof.
Foundry is not a better way to rip one document. It is a way to keep doing it across all of them.
Start with one analyst question such as holdings, footnotes, rates, or lease terms.
Use models to help build and improve routes, not to answer every row on every run.
The winning route runs across the corpus, works in zero-retention environments, and keeps cost and latency down.
Chat works on one file.
Analysts need the answer across all of them.
A single-document answer is useful, but it is still a one-off.
Re-asking a model on every file is slower, more expensive, and harder to trust at scale.
A tournament builds a reusable route so the answers can land in a dataset and compound over time.
One question.
One tournament.
One compounding dataset.
Pick one answer you want back from the corpus: holdings, footnotes, rate terms, lease terms, or another document-bound field.
Several routes try to answer that same question across the same private corpus without forcing a model to do all the work every time.
The winning result lands with its status, its proof, and a route you can run again.
Once the route is good, run it across older files, keep using it on new files, and keep building the dataset.
Start with one question.
End with a dataset.
What are the current holdings?
Run a tournament over filings and schedules to surface row-level holdings.
Which rows have footnotes?
Attach footnote markers to the exact rows analysts care about.
What is the rate and spread?
Break debt rows into reference rate, spread, floor, and coupon.
What are the lease terms?
Pull term, rent, escalators, and allowances from lease documents.