The executives celebrating AI's share of code output and the developers logging post-deployment bugs are not in disagreement about the tools — they are measuring them at different points in the pipeline. Code volume is visible at the generation stage. The cost of that code shows up later, in review cycles that run longer than before and in incidents that land on the people who did not write the promotional slide. That structural offset is why the conversation about AI productivity keeps producing two genuinely different sets of numbers from…