Ingest a case file end to end

Every agent you build on conveyancing documents depends on this layer, and its weaknesses usually surface only when an agent confidently quotes the wrong ground rent. This project builds the layer properly for one realistic, synthetic matter and leaves you with a harness you can point at other documents later. It uses every lesson: the document inventory and triage, parser routing, tables, plans and handwriting, verified extraction, durable fan-out with idempotency and quarantine, re-ingestion with early cutoff, pseudonymisation, and evaluation with a hand-off contract.

Everything runs locally and on your Claude subscription: Claude Code in headless mode (claude -p) is the model inside your pipeline, a local Temporal dev server is the workflow engine, folders are the object store and SQLite holds the tables. Paid parsers and the Claude API are optional stretch comparisons. Expect 8–12 hours.