Assemble a thought

This is the paper's headline result, live. A language model's thought about "the capital of Italy" can be built from three averaged ingredients — a generic base, an "Italy" part, and a "capital-of" part. Write the sum into a single position inside the model while it reads a different question, and it says "Rome" — as often as it says anything correctly at all.

How to drive it. Pick an entity and a question, then press assemble & speak: watch the dot travel the model's map of thoughts (base → +entity → +question), land next to the real measured thought, and read the model's actual, persisted output in the console. Then tick overdose ×4: the same direction pushed too hard leaves the map and the output collapses into token loops — yet forced to choose among the possible answers, the model still picks the right one 80% of the time. Every output shown is a real k=8 generation from the experiments; nothing is simulated.

Full-page: open the scene directly. Prefer the geometry view (re-marking, injection, syncretism)? It's here.