Choose a model for document extraction

Every agent built for the course's conveyancing scenario starts with the same question: which model, at what effort, with what checks? This project answers it properly for one step, extracting structured data from HM Land Registry title registers, and leaves you with a harness you can rerun whenever a new model ships. It uses every module: tokenisation and the transformer (why some strings are fragile), training (why verifiable checks work and where sycophancy and guessing come from), and inference (grounding, caching, effort and cost per completed task).

Everything runs on your Claude subscription through Claude Code in headless mode (claude -p), which reports token usage, cache usage, latency and an API-equivalent cost for each call. Expect 6–10 hours.