We are increasingly delegating operations, decisions and choices to algorithms that affect major areas such as healthcare, policing, employment and media. They are not just tools; they shape society by automating judgement and decision-making. When the Earth Becomes Flat asks: how fragile is the knowledge encoded within such systems, and can that fragility be made visible?
A small open-source language model is trained from random initialisation on a corpus of scientific material, developing a working representation of a spherical Earth. It is then fine-tuned on a concentrated corpus of flat-Earth misinformation: a closed loop in which one signal dominates without contradiction.
The work records how the model's outputs and internal representations change across both phases: it quickly begins producing flat-Earth claims, while much of its earlier internal structure remains. That record becomes this atlas, derived from the model's activation patterns.