Ingest a case file end to end

The pipeline is a build graph with a workflow engine on top. Every stage is a plain function that checks for its artefact under a key made from what went into it: the PDF's hash and page for parsing, the raw text's hash for normalising, and the hash of a document's normalised pages for extraction, each combined with a stage version that hashes the code settings, prompt text, schema and model. The decision that matters most is that models read and code decides. A vision model transcribes scans and a language model proposes field values with quotes, but code checks the quotes against the page, does the table arithmetic, measures the plan's colours, reads the tick boxes, and routes anything uncertain to a person. All documents are synthetic; the people, companies, addresses and title numbers (ZZ6…) are invented.