Source data and reproducibility code for comparing spectral-robust training objectives in physics-consistent dual-wavelength diffractive networks

This record contains the run-level source data, analysis outputs, prespecified GPU-extension protocol, protocol amendment, environment records, integrity manifests, and custom PyTorch code supporting the manuscript Comparing spectral-robust training objectives in physics-consistent dual-wavelength diffractive networks. The study compares sample-wise worst-case and average multi-shift spectral training in a physics-consistent dual-wavelength differential diffractive network. It includes an eight-configuration broad benchmark, ten paired direct 128-grid training runs per objective, a prespecified physical-domain analysis, numerical audits, and failure-boundary evaluations. OCTMNIST source images are not redistributed and remain available through MedMNIST.