Why do simulations introduce correlation artifacts that adversarial training alone cannot detect or eliminate?

I appreciate the thoughtful question, but I need to be direct: the preprints provided don't address simulation artifacts or why adversarial training alone fails to detect them.

Your question asks about a specific failure mode in computational systems—correlations introduced *by the act of simulation itself* that adversarial training can't catch. That's a question about numerical methods, perhaps in physics or machine learning, where the simulation's discretization, approximation scheme, or data generation process introduces spurious statistical structure that robust training procedures can't eliminate.

The archive materials here focus instead on:

• Quantum noise and quantum advantage [AG-2024.07-1643]: how environmental decoherence destroys quantum computational power
• Weak correlations in neural network training dynamics [AG-2024.01-576]: how parameter correlations vanish in the wide-network limit, enabling linear approximations
• Adversarial robustness of classical and quantum classifiers [AG-2024.01-2126, AG-2024.02-1732, AG-2024.04-2186, AG-2024.05-2145, AG-2024.07-1685]: how to attack and defend ML models

None of these directly explain why a *simulation itself* generates artifacts that adversarial training cannot counteract—a question that would more naturally appear in papers on computational validation, generalization bounds, or systematic biases in synthetic data.

If you're asking about adversarial robustness in a different context, I'm happy to explore that; otherwise, this question may fall outside the archive's current scope.