One Readout, Many Terrains: Bilingual LLMs Do Not Share Concept Space at the Decision Level

Ask one bilingual language model the same open-ended question in two languages and it commits to different concepts: "choose a weather condition" settles on cloudy in Japanese and rainy in English, reproducibly. We show that bilingual LLMs do not share a concept space at the level of the committed decision — across 11 contexts, the chosen attractor agrees in only 2, exactly the level chance alone predicts and far from the near-total agreement a shared terrain implies; the temperature at which each decision collapses shows no detectable cross-language correspondence (ρ=+0.002, n=11). What is shared is the readout law: a gap law and a single-pass landscape prediction hold in both languages. We then show the language-specific terrain is a physical object in the residual stream: grafting a single vector transplants the attractor across languages, carrying its depth and surviving full sampled generation, while an agreement control never moves. What is shared is not the concept, but the mechanism that reads it.