Right, there seems to be some assumption that LLMs must respond in a way that its both congruent to the trainer and user but if the user requests responses that

Right, there seems to be some assumption that LLMs must respond in a way that its both congruent to the trainer and user but if the user requests responses that aren't congruent to the trainer, then the model must refuse. I'll just flat out say it - this premise is flawed.