Building worlds that hold together

The failure mode of a generated environment is not that it looks fake. It is that it stops being true about itself the moment somebody looks closely.

A candidate opens the Q3 forecast, notices that the headcount line does not reconcile with the org chart they were sent, and asks the finance lead about it. If the world cannot answer, the exercise is over. The problem is not the bug itself. It is that they have now learned that nothing in front of them is load-bearing, so the rational move is to stop investigating and start performing.

Consistency is a property of the graph, not the documents

So the documents are not the artefact. The artefact is a graph: people, the things they own, what they know, what they are wrong about, and how each of those facts reaches the page. A spreadsheet is a view onto that graph. So is an email, so is the thing the finance lead says when you ask them at eleven o'clock on a Tuesday.

Getting this right meant giving up on generating documents directly. A model asked to write a convincing forecast will write a convincing forecast; a model asked to write the forecast this company would have, given what its finance lead believes and does not yet know, writes something a candidate can actually interrogate.

The people are the hard part

Colleagues in a simulation are not chatbots with a personality prompt. They have a desk, a horizon of what they can plausibly know, and their own view of the work. That means they can be genuinely unhelpful in the way real colleagues are unhelpful: busy, partial, occasionally confident about something that is not so.

That last property is not a defect we tolerate. It is most of the signal.