Track 02 · RetrievalLesson 11 of 11
Contract review and legal work
Review contracts against a playbook, cite every clause, and never invent law.
- Specialisation · Legal
- 35 min read
- 75 min hands-on
- Northvale Bank
- Pro
What you'll be able to do
01- Describe how an in-house legal team works a contract: matters, playbooks, positions and fallbacks, redlines, escalation and sign-off.
- Segment a contract into numbered clauses and classify them using the document's structure as well as its words.
- Encode a playbook as data: required, preferred, fallback and never rules, each with the only authorities a finding under it may cite.
- Produce a deviation report in which every finding carries a pinpoint citation and a verbatim quote, and refuse findings that don't.
- Protect matters (confidentiality, ethical walls, privilege) and evaluate the review against clauses a lawyer labelled.
Before you start
- Answers with receipts: grounding and citations: answer contracts, quote-then-answer and checking attributions after generation. Here, the legal version: clause numbers and verbatim text.
- Parsing and chunking documents that fight back: getting clean text and structure out of Word and PDF files. This lesson starts from parsed text.
- Retrieval that respects who can see what: filtering retrieval by who may see what. Here, by matter.
- Python at the level of reading a few hundred lines. The worked example runs offline.
Starting here? This lesson anchors the legal elective. Two words used throughout: a matter is one piece of legal work with its own file, team and access list (reviewing one supplier's contract is a matter); a playbook is the legal team's standard position on each clause type, what it will accept instead, and what it never accepts.
The short version
An in-house legal team reviews every supplier contract before procurement signs. Most of a first review is mechanical: find each clause, compare it with the playbook, write up what differs. That's where tooling helps, and where it fails if it can't be checked. A lawyer reads a finding, opens the clause it names and compares the words. If the clause doesn't exist, or doesn't say what the finding quotes, the tool has wasted the lawyer's time and lost their trust. If the tool cites law it didn't retrieve, it can do worse: in 2023 a US federal court sanctioned two lawyers and their firm for filing opinions that a chatbot had invented.
So the design starts from citations. Segment the contract into numbered clauses. Classify each clause by type, using the heading it sits under as well as its words. Check each one against the playbook, which is data the legal team owns. Report every deviation with the clause number, a verbatim quote, the playbook rule and its position. Then run every finding, including any a model drafted, through a deterministic check that refuses findings whose clause doesn't exist, whose quote isn't in the clause, or whose authority isn't one the rule lists. The lawyer signs the report; nothing goes to the supplier before that.
- Every finding cites a clause by number, with a verbatim quote. No citation, no finding.
- The playbook is the law the tool knows. A finding may cite only the authorities its rule lists.
- Structure beats keywords. A 72-hour notice under "General" isn't a breach-notice clause.
- Matters are walled. Retrieval, logs and prompts are scoped to the matter, and privilege survives the pipeline.
- The lawyer signs. The tool drafts the first pass; it doesn't negotiate, approve or send.
Pro lesson
The rest of this lesson is in Pro.
The mental model, worked example, field note, failure table, exercise and self-check, with the voice tutor.
- Every stage and specialisation, in full
- Every lab and every role project
- Halden weeks 2–10, on your record
- Every drill, live in the playground
- All solutions, plus the server judge
- Lab Assistant and every Teach with AI