It started as a game no one was meant to play.

Our founder, Patrik Hansson, had long carried an idea for a different kind of civilization game: not one you steer and build, but one you watch — a world growing entirely on its own, with the player forbidden to touch it.

That required an unusual engine, grounded in research on what shapes individual traits, how conflicts arise, what carries across generations, and how much surrounding resources steer a society's path. No scripted events, no predetermined arcs. To study them, the engine runs thousands of simulations across centuries of simulated time, every event replayable exactly.

Then came the question that changed everything: if an AI were put in charge of such a world — would it be better off than if it had been left alone?

The game's core mechanic — watching without touching — was exactly the measuring instrument AI evaluation was missing. The untouched run is the point of comparison: run the same situation again with an AI at the helm, and the difference is the AI's footprint. Not how busy it looked, not how well it did isolated tasks. Whether the system it was responsible for ended up better or worse because it was there — and whether it knew when to leave well enough alone.

Today, that game is a measuring instrument. Across 2,415 runs we recorded no deviations — that is what we observed, not a claim about runs we have not done. We wrote it up in The Steward's Paradox. Being good at tasks and being trustworthy with lasting responsibility are not the same thing, and the failures that matter most never show up in a single run.

AI agents are being handed longer responsibility than ever. We're building the instrument that shows whether they can carry it. No stamp, no verdict — a number you can check yourself.