Most legal AI tools answer questions the way a student might after skimming the first page of a textbook: confidently, but from the wrong sources. LISA was built on the opposite principle. Before it writes a single sentence, it reads the case file, all of it, and then reasons about what it found.
Start with the matter, not the model
Every LISA conversation is bound to a matter. That boundary is not just a folder label; it is the outer perimeter of what the Council is allowed to see. When a lawyer asks a question, the Router Agent first decides whether the task calls for direct analysis or a deeper agentic run. For anything substantive, the run is handed to the Council Agent, which begins by listing every document in the matter vault.
This list is not cosmetic. LISA uses it to build a retrieval plan. It knows which files are pleadings, which are correspondence, which are contracts, and which are precedent drafts because the Vault parser classifies them during upload. The plan then selects the files most likely to contain the answer, ranks pages within those files, and extracts the relevant passages.
Citations before conclusions
Once the text is extracted, the Council structures its reasoning through the framework appropriate to the question, typically IRAC or CREAC for legal analysis, or a direct claim-evidence chain for factual questions. Every substantive claim in the final answer is anchored to a citation. The citation is not a footnote added later; it is the address of the exact source file and, where possible, the clause or paragraph that supports the claim.
The result is an answer you can verify. A partner can click a citation and read the original language. An associate can see why LISA chose one source over another. And because the Council includes a Paralegal Agent that independently checks facts and precedent, the reasoning is cross-examined before it reaches the user.
No model memory, no leakage
LISA does not rely on weights memorised during training. It relies on the documents in front of it, read fresh for every question. That design is what makes matter isolation meaningful: the model never carries facts from one matter into another, and your documents never train anyone's model. Reading the entire case first is not a feature; it is the foundation of everything LISA writes.