An External Bayesian Controller for Environment-Isolated LLM Response Preference

Users steer large language models (LLMs) through instructions, memory and provider settings. These mechanisms do not keep an explicit, updatable estimate of what a user prefers, do not separate why feedback was given, and do not carry consistently across providers or across private and work settings. To show that an external Bayesian controller can maintain and update environment-isolated user preferences independently of LLM weights, and select response strategies probabilistically, expressed as an instruction generated for a target LLM. Delivery of that instruction to a provider, and the provider's compliance with it, are not evaluated.