The model is asked one narrow question per filing. Everything else is plain code: which pages it reads, which label the answers lead to, and whether that label is safe to publish. Select any step to see what it does and where it lives.
Code, no modelModel callPublishedHeld for a person
01IntakeEvery new filing
02ReadChoose what the model sees
03AskFacts, not a label
up to 2 looks
no
yes
04DecideRule order in code
05CheckNothing ships unchecked
agrees
disagrees
One filing, call by call
What actually goes to the model and comes back, for one real filing: Patel Engineering's 204-page annual report. It is one of the few filings where the model used its page tool, so it shows the whole loop. Open any part to read it in full.
Loading the trace…
How it is measured
The same pipeline runs three ways. Hand labels say whether it is right. A frontier model on the same filings says how much accuracy the cheaper model gives up. Every call is metered, so the bill comes from usage, not a guess.
