Notes from building cited covenant extraction: how models compare on real agreements, what OCR gets wrong on scans, why the parser runs before the model, how to measure a citation, and how much human review the numbers need. The figures that recur: 97.4 percent citation-exact on the first pass, 99.7 after the second read, and every affected test recomputed inside nine seconds of data landing.
We ran four classes of model over 60 agreements and 1,840 cited fields. The frontier still leads on raw reading, but the gap sits almost entirely in definition chains and citations. With the deterministic pre-pass the production first pass reaches 97.4 percent, and the second read 99.7.
Scanned execution versions, two-column schedules, footnoted thresholds and stamped pages. Nine practices that moved our first-pass field accuracy on scans from 88 to 96.9 percent, with the second read closing the rest, and the one metric that hides the damage.
A parser builds the section tree, the definitions index and the cross-reference graph before any model sees the document. The model then answers from candidates and must cite a span the validator can check. Why the order matters, the failure it prevents, and how it recomputes every affected test inside nine seconds of data landing.
Field accuracy is the metric everyone reports. It misses the quadrant that fails an audit: the value is correct and the clause it points to is not. That quadrant was 4.6 percent of our fields before the validator and is 1.1 today, inside a 97.4 percent first pass, and it is invisible to a spot check.
A leverage covenant is one line. Resolving it means following Consolidated EBITDA into Consolidated Net Income and onward, a median of four hops and up to nine. Where the chains go, how often they leave the document, and what that does to extraction.
Reviewing every field is a second extraction. Reviewing none is a bet. Routing by calibrated confidence, a second read on 15 percent of fields takes our first pass from 97.4 to 99.7 percent. How to set the budget and what to do with the rest.