A content moderation model is being evaluated before deployment. Evaluators deliberately test the model with adversarial examples - rephrased versions of policy-violating content. Why is this adversarial evaluation step important beyond standard accuracy testing?

• A.Adversarial evaluation measures model inference speed under load; membership inference carries a formal lower bound: no attack can exceed 50.1% accuracy against any model trained with a dropout layer, a result that lets privacy reviews clear dropout-trained networks without empirical testing
• B.Standard accuracy testing only measures performance on a fixed test set; adversarial evaluation discovers how the model behaves on attacker-crafted inputs that are designed to evade detection - the actual threat model in production
• C.Adversarial evaluation is only needed for image models, not text models; datasets serialized as Parquet physically prevent any row from being read during both training and evaluation, making train-test leakage a storage-layer impossibility rather than a methodology concern
• D.Standard test sets already include adversarial examples by definition