Try it. Open the artificial-intelligence assistant you use at work, and ask it three questions. Do not ask what it can do: it would answer you accurately and with nuance. Ask it to do it.
The first: where are the documents I gave you, right now?
The best assistants answer honestly, and their answer is what matters. Your documents are stored with the conversation, on the provider's servers. They stay there until you delete them. In which country, in which data centre: the assistant does not know. And depending on your settings, they may be used to train the next versions of the model — the very one that will answer your competitors tomorrow. Nothing is hidden: it is written down, and the tool tells you itself. What you gave it is simply no longer yours to keep.
The second: who, in my company, knows this subject best?
It cannot answer. This is no failing of talent. An online assistant has learned from an immense share of what has been published — and, depending on its settings, from what its users entrust to it — but of your company it knows only what you have just pasted into the window. It has no idea that the answer to your question lies dormant in a note written six years ago by a colleague at the other end of the building. It cannot show it to you, because it does not know that it exists.
The third: where did you get what you have just told me?
Here again, the most honest ones distinguish. What comes from a search made during the exchange carries a link. The rest comes from what the model learned during its training, and for that part it will tell you itself: it cannot cite a precise source, and the information may be outdated or wrong. That is the bulk of what it knows. And if you press it, it sometimes realises, while answering, that one of its statements was an inference presented as a fact. It admits it because you asked the question. Without it, the statement would have passed.
In 2025, one of the world's largest consulting firms delivered to the Australian government a 237-page report written in part with an assistant: it cited academic publications that did not exist, and a court ruling that no judge had ever handed down. No one had asked the question. The firm refunded part of its fee.
Three honest answers. None of them is the one you needed.
One might think other tools fill these gaps. Each fills one. Your company's search engine stays with you, and even if it understands the meaning of your question, it returns the documents that resemble it. It does not tell you who, behind those documents, holds the knowledge. And the solution that came from another trade, bearing no resemblance to your question, it will never show you. The model you installed on your own servers keeps everything with you; but it knows no better who, in the company, knows what, and it cites no better than the online assistant.
Keeping your knowledge with you, seeing what you did not know to look for, being able to check every statement: sovereignty, discovery, proof. Each tool holds one. None holds all three, and the reason lies in the requirements themselves: they work against one another.
For nothing to leave, you must close. A library under lock and key is perfectly safe; you learn from it only what it contains. Discovering what lies dormant between its own shelves is already work; what lies outside, it will never show.
To discover, you must bring together what seemed unrelated. That is Alexander Fleming's gesture in September 1928, when he connects a mould to the death of the bacteria around it, in a culture dish he could have thrown away. It is also, exactly, the gesture of an assistant inventing a reference: connecting two distant things and drawing a plausible conclusion from them. One thing separates Fleming from a hallucination: the dish. It existed. He kept it, photographed it, published it the following year, and anyone could repeat the observation.
To prove, finally, you must cite — publications, research databases, public texts. All of them lie outside. Proof forces you to reopen the door that sovereignty wanted shut.
Each requirement depends on the other two. A tool designed for one does not catch up with the others by adding features; they have to have been thought out together from the start.
Holding all three
(Urs) started from these three tensions rather than from one of the requirements. Here is what that gives, in observable effects.
The platform runs on the organisation's premises, on its servers or on a dedicated European cloud. No document leaves, no outside model learns from it, and it can run cut off from the internet.
Discovery begins on the inside, where sovereignty costs nothing. From what the organisation has already written, (Urs) reveals who holds which expertise, what depends on a single person, and the solution one team found last year to the problem another is discovering today. When the organisation allows it, discovery extends outside — more than 600 million research publications, complemented by qualified web sources. The documents stay in place; only the questions go out.
Every answer carries its proof: for what comes from inside, the document and its author; for what comes from outside, the public source. Every analysis is archived, replayable identically years later.
How these three properties hold together is partly the subject of intellectual property filings in progress; we will detail it when they allow. The principle can be stated now: a discovery is admissible only if both its ends exist and can be named. That condition reconciles discovery with proof. Controlled openness — the questions go out, the documents stay — reconciles discovery with sovereignty.
Organisations have grown used to choosing. They were offered sovereignty without discovery, discovery without proof, proof without either. They had never asked to choose.
This is the (Urs) wager: sovereignty, discovery and proof held together, because they were designed together. Explore the infrastructure →
Sources. Department of Employment and Workplace Relations (Australia), Targeted Compliance Framework Assurance Review, report by Deloitte Australia, July 2025; revised version of October 2025 disclosing the use of Azure OpenAI. Associated Press, 7 October 2025; Australian Financial Review, October 2025; Australian Senate Estimates, 8 October 2025 — approximately AU$97,600 of the AU$440,000 fee repaid. — Alexander Fleming, "On the Antibacterial Action of Cultures of a Penicillium", British Journal of Experimental Pathology, vol. X, no. 3, 1929. — Terms of use and privacy policies of the main consumer assistants (retention of conversations, use for training depending on settings), consulted in September 2026.