Notes from the build

Most legal-AI demos answer a question and ask you to trust the answer. The peer-reviewed evidence says you should not: even legal-specific tools have been caught inventing citations a meaningful share of the time, and models are not reliable at checking their own work.[4]

So we built the workspace around a different idea: don't trust, verify. An independent model checks the reasoning, and whether each citation is real and verbatim is confirmed in code against the official source. That check sits outside the model, so it does not move when the model does — the workspace gets better as the models improve, without ever loosening the guarantee.[6]

This is the first of a series of build notes. Future posts will go deeper on the architecture, the evaluation method, and what we learn from the recorded case studies.