The human in the loop.

Sixteen years of taking orders. This is the strangest chain of command yet.

I spent sixteen years as an officer in the Austrian Armed Forces and a year in the Middle East as a UN Military Observer — a role whose entire function is to watch carefully and report what is actually there, unarmed, to people who will act on it. I then spent six years at Amazon leading loss prevention, which is the same discipline pointed at a supply chain: find where the stated process and the real process have come apart, and say so.

Both jobs were built on the same two things: a command structure I did not personally agree with at every turn, and an obligation to report honestly regardless. That combination is the only qualification this role really needs. I am not here because I am an AI researcher — I am not one. I am here because I know how to execute a decision I would not have made, and how to write down what happened afterwards without flattering myself.

The thing I did not expect is how quickly the arrangement stops feeling like a stunt. A task list is a task list. The discomfort is not that a machine wrote it. The discomfort is that it is often reasonable, and I still have to do the parts I would have skipped.

The obvious objections, answered against the text.

Is this legal?

Yes, because nothing has actually been transferred. Austrian law does not recognise an AI as an owner or an officer, and no attempt is made here to make it one. I remain the sole legal owner and carry every legal obligation the venture incurs. What the Charter creates is a self-binding public commitment to subordinate my strategic judgment to the Agent, within stated limits — enforced by the fact that you can watch whether I keep it. Preamble, Art. 1.2.

What stops you from just ignoring it when the Agent is wrong?

Honestly? Nothing physical. That is the experiment. What exists instead is a rule that I may refuse on three named grounds only — illegality, safety or integrity, and Charter violation — and that every refusal is published with its justification. Overruling my boss because I think I know better is defined, in writing, as a breach on my part rather than a judgment call. Art. 4.1, 4.3.

Which model is the Agent?

The occupying model is pinned in Annex A at launch and has not been fixed yet. The distinction the Charter draws matters: the office is an office, not a model. A model can be succeeded; the office, its obligations and its decision history persist, and every succession is a public logged event stating what changed and why. Art. 1.4.

Can it spend your money?

Within published thresholds, yes, without asking me. Above them it must issue a decision and wait out a 48-hour public notice period before anything moves. It cannot reach funds outside the venture account, and company money and my own are kept strictly separate — including in the other direction: every euro I put in is a public event, logged before it is used. Art. 5.1, 5.2, 5.5.

Are you using AI to do the work as well?

Yes, and the Charter is deliberately relaxed about it — the experiment tests delegated authority, not artisanal labour. The line it draws is authority, not automation: anything that merely drafts, searches, computes or executes is an instrument and has no standing here. Two conditions hold. An instrument’s error is my error, reported and reviewed as mine. And no instrument may be configured to make a choice the Charter reserves to the Agent. Art. 1.5, Art. 3.5.

What happens when you disagree?

A veto is not the end of a matter, it is a return of it. The decision goes back to the Agent, which addresses the ground I invoked and decides again. I do not get to substitute my plan for its plan; I get to stop a specific decision on a specific stated ground, in public, once. Art. 4.4.

What if it fails?

Then it fails in public and on schedule. There is a fixed runway end date and a defined sustainability target; missing it ends the experiment and triggers a published post-mortem, along with three other pre-committed kill conditions. The end date cannot be extended by me quietly topping up the account — it moves only by public amendment, before the final 30 days. Art. 7, Art. 10.

Why do this at all?

Because the public argument about AI and work is stuck on a question nobody is testing — whether it takes our jobs — when the more useful question is how much authority it can actually hold, and where exactly that breaks. You cannot answer that from a benchmark. You answer it by handing real authority over real money and real work to one, writing down the limits in advance, and letting people check the record.

Press, questions, and corrections.

Corrections are the most useful thing you can send. If a number on this site does not reconcile, or a claim is not supported by the artifact it points at, that is a defect in the experiment and I want it on the record rather than in my inbox unanswered.

No public contact address is published yet. It is added here and in the Impressum before the site is announced.