On-premises document analysis · UK

ARTEMIS NODE

A document analysis appliance that sits in your office, reads your clients' papers, and has no route off the premises. No cloud account, no data leaving the building, no external connection of any kind — and it does not stop at five o'clock.

1. Nothing leaves the building

The appliance is a sealed, headless unit connected to your office network and nothing else. It holds no cloud credentials, makes no outbound connections and sends no telemetry. Your clients' documents are read from your storage, processed locally, and written back to your storage.

2. Every quotation is checked against the source document

Language models fabricate. The usually-quoted answer is “use a better model”; the reliable answer is to check the arithmetic. Every quotation the appliance produces is mechanically matched back against the text of the document it came from, and the report says, per quotation, whether it was found. Where it was not, the report says that too.

Analysis is produced by two independent passes that read the document separately. Where they disagree, the report shows the disagreement rather than averaging it away. A third pass then interrogates the first, looking for what it missed or overstated.

3. What it does

• Clause-cited review. A document is read and the findings are delivered with the passage each one came from, so the analysis can be checked against the source rather than taken on trust.
• Chronology and key dates. Every dated provision, extracted twice independently and checked against the document, laid out in order — including the entries the two passes disagreed on.
• Redaction. Identifiers are found by pattern and, for most of them, confirmed by check-digit arithmetic — the same arithmetic that validates an NHS number or a National Insurance reference. A redacted copy is produced, and then re-scanned to confirm the redaction worked. This is the feature we would show you first, because its result can be verified rather than believed.
• Scanned paper. Documents without a text layer are read by offline optical recognition. Court bundles, signed pages and deeds are usable input rather than something to be retyped first.
• Disclosure indexing. Documents are numbered, each one with a note of how it was read, and a chronology in which every entry cites the document it came from.
• Attendance notes. A recording is transcribed on the machine and drafted into a first-draft note for the fee earner to correct. The transcript is delivered alongside it, with timestamps, so every figure in the note can be checked against the recording.

4. Every report states how it was made

Each output carries a record of its own production: the version of the appliance, when the run happened, a hash of every document as it was received, and which model produced the result — identified by the hash of the model's own bytes, not by its name, because two files with the same name can be different models. The record also states, as a required field, what the appliance did not do.

Where a document cannot be read, or a recording cannot be transcribed, the appliance refuses and says why. It does not return an empty result that reads like a clean bill of health.

5. It works when you are not there

The appliance sits in your office and does not stop at five o'clock. A bundle left with it overnight is done by morning, and the report is waiting when you get in.

It does not tire, and it does not skip. A person reading two hundred pages at the end of a long day is not the same reader who started them; the appliance is. That is not the same as being right — which is why every finding cites its source — but it is the part of the work that can be taken off a fee earner without taking the judgement with it.

6. How it is delivered

Every node is built to order on our bench and configured for the firm that ordered it. There is no generic image sitting on a shelf, and nothing about your setup is a default we forgot to change. We install and commission it with you; your IT provider gives it a static address on your network, and that is the extent of their involvement.

On your network the appliance serves one port to your intranet and nothing else. Reports are written to storage you control, and the record of how each one was produced goes with it.

7. What we cannot do — which is the point

We do not run demonstrations over the internet. The appliance has no route off your premises, so neither do we. Anything we showed you remotely would have to come from somewhere other than the machine you are considering buying.

So instead: send us a bundle you already have, and we will put it through our own baseline node and give you the output unedited — the clause analysis, the chronology, the redaction report, and the record of how each was produced. Read it the way you would read a colleague's work. If it is not good enough, you will know before you have spent anything.