Supervised Finetuning Sets Cross-Language Representation Geometry; LLM Judge Scores Fall Further Off the Axis

Users of language models in languages other than English routinely report that assistant output reads like translated English, and the stage usually blamed is RLHF. Using the fully open post-training ladders of Tülu 3 (Llama-3.1-8B) and OLMo 2 7B (base → SFT → DPO → RLVR), we measure where the cross-language representation change enters and what geometry it has, using a base-anchored cross-language register-separation axis estimated from content-matched Chinese–English pairs, output-side stylometrics, blind LLM-judge naturalness ratings, and two native raters. The mid-band convergence of Chinese and English on this axis is almost entirely a supervised finetuning effect. The Chinese–English gap contracts at SFT (by 12.8% on Tülu 3 and 20.0% on OLMo 2) and then moves by under 4% of that contraction through DPO and RLVR, whose increments have cosine at most 0.23 with the total drift; DPO contracts the gap further only in the last three layers. On the multilingual-capable Llama base the two languages converge symmetrically; on the English-centric OLMo base the Chinese side does the moving, and only the OLMo contraction survives orthogonalization against a content-disjoint language-identity axis. After SFT, two LLM judges from different vendors score Chinese chat output lower (significantly on OLMo 2 under both, on Tülu 3 under one), while list formatting and Latin script become more frequent and account for 25–52% of the paired decline. English scores stay at ceiling. Two native raters of 50 paired prompts disagree with each other (ρ = 0.08) and neither shows a significant change, so a loss of human-perceived naturalness is not established. An inference-time subtraction of the SFT-set drift direction restores 42% of the representation gap without lowering judged usefulness or English output, and same-norm random directions restore none, yet judged naturalness does not recover. Every representation-level signature replicates in Japanese, whose chat output a single LLM judge scores below Chinese at every instruct stage. Code, prompt pools, judge prompts, human ratings, and analysis scripts are in the supplementary archive.