<p>Extreme class imbalance and the subtlety of rare fraudulent patterns hinder fraud detection in financial domains. Most existing methods struggle to balance sensitivity and specificity, often overfitting to noisy minority examples, failing to generalize across ambiguous decision boundaries, or ignoring instance-level uncertainty. We propose AURNA, a novel neural network framework that integrates the Pinball loss function to enable quantile-aware, instance-specific penalization. Unlike conventional loss functions, AURNA dynamically adjusts penalties based on prediction confidence and class membership, enforcing quadratic penalties for uncertain or misclassified cases, while linear penalties are applied for confident predictions. This approach yields adaptive decision margins and mitigates overfitting to easy examples, making the model uncertainty-centric and thus more generalizable. We also introduce a cost-sensitive quantile selection scheme that balances sensitivity and specificity, taking into account the economic impact of false positives and negatives, to further optimize performance. Empirical results on a real-world car insurance fraud dataset with a 6% fraud rate demonstrate that AURNA achieves a 96% sensitivity while maintaining overall performance and reducing fraud-associated costs. These findings underscore the promise of granular, margin-aware penalization in high-stakes imbalanced classification.</p>