A software licence is rarely one document.

An answer is only as good as the clauses the model gets to read. One agreement fits in a context window; a family of master terms, orders, schedules, amendments and DPAs may not — and as the pile of related documents grows, finding the handful of provisions that actually control a question becomes the real work.

Handing the whole family to a frontier model works for a small stack. The costs arrive as it grows: every document in the prompt is confidential text leaving your control, and a complex enough family will not fit the window at all. Either way the model reads a subset — the question is whether anything chose it.

AgreementAtlas builds that context deliberately: a knowledge graph of every clause, definition, amendment and precedence relationship, assembled before any model runs. Retrieval walks the graph — which definition controls, which order amends what, which document wins a conflict — and hands the model only the provisions that govern, with the exact source text attached.

Context that good changes what the model needs to be. A small open-weight model running locally on a MacBook Pro M5 Pro competes with frontier cloud models on these questions — with confidential agreement text never leaving the machine, and every conclusion tied to the exact source clause.

How the graph is actually built — and what it does not do yet →