Every precise instrument is first calibrated: you feed it a signal of known strength and observe the response. The digital city of Montelago was treated the same way. Researchers created 120 virtual characters, each given a personality: hidden trust in municipal authorities, assessment of news credibility, and perceived appropriateness of information. These numbers were known only to the creators.

Then a language model (a program that generates text, like an extremely smart autocomplete), built on learning circuits with a special architecture and ability to focus on what matters, and trained on vast amounts of text, began generating character responses to news. It wasn't given the hidden numbers: it only saw their biographies. The news varied: from a water pipe repair to accidents and murky water. They added a control message about a city festival — harmless, like a placebo. This entire numerical experiment was repeated several times, varying the generation temperature — a setting that adds randomness to responses (the higher it is, the more different the answers could be to the same question).

Seven checks showed that hidden trust and responses matched almost perfectly (a correlation of 0.9). Without news, assessments didn't change. The instrument noise (variation across repeated questions) was small, like the trembling of a scale's needle within one division. But one news item, meant to slightly improve mood, unexpectedly worsened it. This wasn't a model malfunction; it was a flaw in the message itself: the text inadvertently highlighted problems. After correction, everything fell into place. So calibration helped improve not the model, but what we say to it.

This approach turns digital residents from a toy into a genuine measuring tool. Now, when preparing important announcements (say, about contaminated soil or road repair plans) officials can first test reactions on a virtual city, just as scientists check their instruments. Montelago is calibrated and ready for real measurements.