Accurate Surrogates Can Hide Exact Response Aliasing: A Post-Training Audit for Calibration

Scientific surrogates are commonly selected by held-out error, but inverse problems also depend on variation between simulated parameter values. We introduce a post-training audit that identifies response directions unconstrained by those values and tests their downstream effect. Applied to a hybrid polynomial--MLP surrogate trained on 5,248 photonic-crystal FDTD simulations, the audit reveals an exact two-dimensional slab-thickness alias. Cubic and quartic feature-map terms are indistinguishable from linear and quadratic terms at the three simulated thicknesses. The median maximum deviation from the quadratic through those predictions is 5.35\,nm. A degree-constrained retrain reduces this to 0.10\,nm with only a 0.3\% wavelength-RMSE change. Responses identical at all simulated thicknesses produce different calibration objectives. Six SVI runs under the deployed response infer thicknesses spanning 1.83 percentage points despite similar losses and held-out errors. Thus, held-out accuracy can miss surrogate-induced response non-identifiability, and off-grid simulation or metrology is required to resolve the physical response.