You have 0 free articles left this month.

What a fact layer does on a live matter

A fact layer reads every document in a matter and holds what they establish, where they conflict and what is missing. Mary sits between your documents and your AI.

November 17, 2026 • By Mary Technology
Share this article on:
expand image

A fact layer is the part of a legal AI setup that holds the facts of a matter. It reads every document on the file, records what each one establishes, marks where two documents conflict, and notes what the file refers to but does not contain. Above it, the reasoning layer argues and drafts. The fact layer is what that reasoning works from.

In an Australian firm the documents are rarely in one place. A disputes matter usually draws on a document management system such as iManage or NetDocuments, a practice management system such as Smokeball, a shared drive, and the email of whoever has been on the file longest. A fact layer has to read across all of them, because a fact that sits in only one of them is still a fact in the matter.

The clearest test is the next tranche of discovery. Read on its own, a new tranche is compared with the last one from memory. Read into a fact layer, it joins the same matter as everything before it. Where the new account conflicts with the one already on the file, Mary flags the clash instead of leaving it for someone to notice.

The second test is handover. Disputes run for years and the people on them change. When a solicitor picks up a file they did not open, the fact layer is what they inherit. Every document named and dated, the chronology already built, the open questions already marked, and each fact pointing back to the page it came from. The alternative is reading the boxes again.

The third test is disagreement. Two witnesses describe the same meeting differently. An invoice and a file note put the same payment on different dates. A fact layer keeps both accounts and marks the conflict instead of choosing between them. The same goes for absence. A document three witnesses describe and nobody has produced is recorded as missing, because on a dispute that is often the most useful thing on the file.

None of this is a large-firm problem. A two-partner disputes practice and a national firm both run matters where one document changes the advice. The national firm has more people to catch it by hand. That control is expensive, and it is the first thing to come under pressure when the matter gets busy.

Once the facts are held in one place, the AI a firm has already bought can work from them. Mary answers Claude, ChatGPT and Harvey. A lawyer asks the question where they already ask it. The answer comes back from the matter, not from whatever was dragged into the window that morning. Nobody has to learn a new place to work.

A fact layer has a limit worth naming. Deciding whether something is established or merely asserted is legal work. Software can put the conflict in front of a lawyer, show the page it came from, and record who checked it. It cannot decide what the conflict means, and it should not try.

Mary is built to be that layer. As of November 2026 Mary reads from iManage, NetDocuments and Smokeball, and connects to Claude, ChatGPT and Harvey. Documents come across from the system they already sit in, and Mary does the reading. The firm's own permissions carry through, so a connected tool sees only the matters the signed-in lawyer can already see. Sign-off happens in Mary, not in the tool that asked the question.

The reasoning layer will keep improving, and firms should keep buying the better one. The facts underneath it only have to be built once, and every model that arrives after this one will work from them. The place to test that is the matter on your desk now.

Book a demo


Mary Technology builds precise AI for litigation and disputes teams. It reads every document in the matter, keeps contradictions and gaps visible, and connects to the AI tools a firm already uses.