You own the machine, the model, and the record of every decision it touched. Privileged material, protected health information, trade secrets, and client files never leave the building — and you can prove exactly how the work got done.
Every AI arrangement a firm can buy today is access to someone else's machine, running someone else's model, under terms that can change. This is the other kind.
It's in your building, in your cabinet, on your power. Nobody can raise its price, deprecate it, rate-limit it, or turn it off.
Open-weight models under Apache 2.0, pinned to a version in the build recipe. The model that answered in March is the model you can point to in March.
Every consequential release is sealed and stays with you. It doesn't live in a vendor's log, and it doesn't disappear if we do.
The machine is sized to the firm at assessment. The software, the review step and the seal are the same on every size.
M5 Pro with 64 GB of unified memory. Runs Gemma 4 for drafting, routing and reading scanned documents.
M5 Max with 128 GB of unified memory and 614 GB/s of bandwidth. Runs gpt-oss-120b alongside Gemma 4.
128 GB of unified memory on the NVIDIA software stack. One serves a firm of twenty-five; two, clustered over a 200 Gb/s link, handle considerably more.
The upgrade path is a second box, not a server room.
A question comes in and routes to the vertical that governs its domain. It answers from documents your firm owns — which is why it can't cite a case that doesn't exist. The model isn't recalling a citation from training. It's retrieving a document that's actually in your library.
A small fast model classifies which briefcase governs the question. The classification is itself a judgment, so it's recorded — including when it's uncertain.
A local index returns the relevant passages from documents your firm owns, tagged at ingest with where each one came from.
The answer is written with the source documents displayed alongside it. Nothing has left the building at any point.
Nothing consequential is released until a named human reads it. Not a checkbox — a person whose name goes on the result.
The signature captures who reviewed, when, how long they held it, and what was in front of them when they decided.
LedgerGuard binds the document, its sources, the model, the signer and the moment into one verifiable record.
Everything runs on the node by default. A frontier model is used only by the firm's written choice, per matter, and that choice is recorded with the result.
The seal isn't a logo on a PDF. It's a cryptographic hash over everything that mattered at the moment of release — and whoever receives it can verify it with standard tools, without asking us for anything.
Delivered and installed in weeks, not months. Professionally provisioned and operated — not sold and abandoned.
The build is assembled, your library loaded, and the whole system verified before it leaves. The golden artifact is a hashed, versioned build recipe.
Before first activation, the deployment model is disclosed to your professional liability carrier. Ahead of installation, not after.
A technician racks it and connects power and local network. Physical access only — the installer never has logical access to the system or your library.
Retrieval is tested on your own material at commissioning. That result is part of the record, not a marketing claim.
An operator runs it inside the firm, learns how the practice actually works, and builds the library around it — until your people can run it themselves.
We should talk about your firm — what it handles, what it can't afford to leak, and what you'd need to show if someone asked.
Start the conversation