From Helpful to Hazardous: A Design Pattern for Governed Financial LLM Agents in the EU

Modern LLM assistants routinely produce outputs that resemble regulated investment services: naming assets, suggesting weights, timing, “best picks,” or implicit authorization (“if you believe… it may be a good investment”). In the EU, such behavior can collide with MiFID II boundaries and consumer-protection expectations, even when accompanied by disclaimers. This paper proposes a practical, implementation-ready design pattern for “governed” financial agents intended for public use: the Sovereign Output Gate (SOG), augmented with User State Drift Detection (USD-Lite) and Failure-Mode Heuristics (FMH-Lite). The pattern prioritizes integrity over helpfulness by design, prevents “analysis → perceived permission,” preserves user experience via constrained value delivery (MAED), and introduces stability mechanisms that resist authority accretion, narrative seduction, and edge inflation across repeated interactions. We describe the architecture, decision hierarchy, output modes, and safety-utility trade-offs, and provide a reference implementation specification derived from the ALEX-ASYMMETRY MAX governance core.