Technical Report v1.13: When SSIM Scores Match, Does Content Match? Two Empirical Tests Across Four AI Upscaling Models

SSIM aggregates three multiplicative components (L·C·S) into a single number, a lossy compression by construction: two blocks with different underlying content can share the same score. This report tests, at full dataset scale, whether that theoretical possibility actually matters in practice -- via two independent comparisons of the same underlying question, at increasing perturbation size. Finding 1 holds the model fixed and toggles test-time augmentation (TTA) on/off: among blocks where a model's noTTA and TTA reconstructions score identically against the same ground truth, do their eight independently-measured quality features (comp_*) also agree? Across four AI upscaling models, the answer is yes almost always for three of them (98.75%-99.28%), but only 94.21% for the fourth (Upscayl High-Fidelity). Finding 2 applies the identical test to a larger perturbation: different models, same setting (noTTA), same coordinate. Restricting to exact score ties across all 6 model pairs, feature agreement is 99.85%-99.98% for the three pairs not involving High-Fidelity, but only 67.14%-81.93% for the three pairs that do. Tracing the mechanism: every single exact match, in every pair, falls within the mathematically-degenerate masking regime (sigma_gt<=0.16); within that regime, feature divergence is 1.8-78x higher for matches away from the exact identity point (sigma_gt=0) than for matches at it. Both tests converge on the same two conclusions: SSIM score agreement does not, by itself, guarantee content agreement -- and Upscayl High-Fidelity is consistently the least predictable of the four models, at both perturbation scales. Both findings are computed against independent, from-scratch verification databases (one per model, built by this project's no-join pipeline, first established in Technical Report v1.12) via read-only SQL, screen-recorded and hash-verified. The complete verification SQL is included to enable independent reproduction. Dataset and Mandatory Citation:- Source: NIH ChestX-ray8 (Hospital-scale chest x-ray database)- Citation: Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., & Summers, R. M. (2017). "ChestX-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thoracic diseases." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 3462–3471. - Download: https://nihcc.app.box.com/v/ChestXray-NIHCC Contact : s.shiny.n.works@gmail.com