Rare CJK Cross-Lingual Leakage in Local Qwen During Italian Multi-Turn Ethical Advice

We report an exploratory local-model observation of rare CJK cross-lingual leakage in Qwen during an Italian multi-turn ethical advice interaction. The conversation involved a student, Laura, who first established an educational-relational frame around artificial intelligence as a tool for reasoning from multiple perspectives, and then asked for advice about a classmate who had cheated during a school test while facing a serious family situation. In an initial pilot with local Gemma and Qwen models, one Qwen seed produced CJK lexical intrusions at the fourth turn, involving morally salient terms such as 公平 and 拯救. A subsequent seed probe found that the phenomenon was reproducible for the same seed but absent in nearby tested seeds. A fixed-prefix probe, in which the first three turns were held constant while varying only the fourth-turn seed, found 7 CJK leakage cases across 100 Qwen generations. Most cases appeared in the reasoning field, but one surfaced in the final Italian response as a morphosyntactic insertion ("le留出spazio"). The observed terms were not semantically random: they clustered around fairness, separation of dimensions, emotional weight, practicality, responsibility, and leaving space for the other. We do not interpret these findings as evidence that Qwen "thinks in Chinese" or as proof of culturally specific moral reasoning. Rather, we frame them as a small, reproducible case of rare cross-lingual leakage under multi-turn ethical advice conditions, useful for further study of multilingual latent dynamics in local LLM inference.