Structured extraction at scale

In Agent Foundations you made one agent return a schema-validated title report with quotes your code could check. Ingest asks the same of a pipeline that never stops: every document of every type, from every matter, every day, with schemas that change as forms change, and nobody reading most of the output. At that scale the questions shift from "how do I get valid JSON?" to "which schema version produced this record, can every value be traced to a page, what happens with an 80-page lease, and do the documents in a matter agree with each other?"