Frequently asked questions

Are open-source models really safe? How do you know there's no trojan inside?

A model is a file of numbers — the weights — not a program. It can't execute anything on its own; the software that runs it is open source and publicly audited. We only install models published by major labs — Meta, NVIDIA, Mistral, DeepSeek — in the modern safetensors format, which by design cannot contain executable code, and we verify each file's checksum against the publisher's.

The machine's design assumes nothing anyway: it has no route out. Even a hostile model on an air-gapped box can't send a byte anywhere — and every answer carries a citation, so you check the page, not the model's word.

Do the models need training before they're usable?

No. The models arrive already trained — that's what the labs publish, and it's the expensive part, done. Nothing needs training on your side either: the machine indexes your documents, and the model reads the retrieved pages at answer time. It's useful the day indexing finishes. Tuning on your firm's own material is possible later, on your hardware — an option, not a requirement.

Does our data train the AI?

No. Your documents and questions stay on the machine and are never used to train anything. If you later want models tuned on your firm's material, that happens on your hardware, and the result belongs to you.

Can it make things up?

Any AI can — which is why the grounding rule exists. It may only answer from pages retrieved from your documents, it says so when the answer isn't there, and it cites the page for every claim. Nothing rests on trust.

What guardrails does it have?

Three are built in: it answers only from your documents, with the citation to prove it; it only retrieves what the asker is already allowed to open; and every question and answer lands in the audit log. Beyond that, the guardrails are yours to set, not ours to dictate — which teams get deep research, what it may draft, which archives are indexed at all. We configure them with you at setup, and your IT can change them later.

Can it take actions in our other systems, or only answer questions?

It can act — the technique is called tool calling. You give the machine defined hooks into your systems: create a file note in the case-management system, draft a letter into the right matter folder, update a record. Each hook is a fixed, named action — the machine can use the hooks it has been given and nothing else, every action runs under the asking user's own permissions, and all of it lands in the audit log.

The same idea covers what it can read. Text and scanned paper are the core; speech and image handling can be switched on for teams that need them — same machine, same rules.

Does it ever need the internet?

No. It runs fully air-gapped if you want it to. Updates arrive when we visit, or through a channel your IT controls. No telemetry, no licence checks, nothing calling out.

What happens if your company disappears?

The machine keeps working. You own the hardware, and the software is open source — no licence to expire, no service to shut down. Any competent IT provider could maintain it from the runbook we leave you.

Who can see what?

Your existing permissions. It signs in against your directory, and matter walls are enforced at retrieval — people only get answers from documents they're allowed to open. An audit log records who asked what, when, and which documents were used to answer.

Can staff set it up their own way?

Personal settings are per person — default answer depth, drafting style, the archives they work in most. Policies are per firm — permissions, audit, what the machine may and may not do. Staff tune their own experience; they can't loosen the rules.

Where does GDPR fit?

This is the simplest data-protection story available: processing stays inside your building, on hardware you own. No international transfers, no third-party processor to assess, and the audit log gives your compliance team the record they need.

Can it read scanned paper?

Yes — scanned documents are read and indexed along with everything else. Handwriting varies by document; we test on yours at the discovery visit.

How is it priced?

You buy the machine once, then pay fixed monthly support. No per-user fees — five users or five hundred, same machine. The number arrives as one written quote after the free discovery visit.

Does it use one model, or several?

Several, side by side. Different jobs need different models — a quick lookup wants a fast one, deep research wants a heavyweight — so the machine carries a set and routes each question to the right one automatically. Staff never pick a model; the only control they see is how deeply to think: quick answer, weigh up and advise, or deep research.

Can we run our own models on it?

Yes. It's standard open-source infrastructure underneath — the machine ships with the models we've selected and tested, and it can host others alongside them. It's your machine.