nx_image_fidelity.nx

SUPERSEDED-FOR-GENERATION (Critic, Blau-Michaeli): this measures DISTORTION only (MSE/PSNR), which is correct for lossless RECONSTRUCTION but MISLEADING for generative quality -- a blurry posterior-mean scores high PSNR yet looks fake. For generation, pair this with the PERCEPTION axis in nx_rd_perception (distribution divergence). Distortion and perception trade off; never judge a generative output by PSNR alone. nx_image_fidelity.nx -- "rebuild it EXACTLY": the measurement gate for image reconstruction, bits-up. You cannot claim a rebuild is exact without measuring exactness; the precision research flagged measurement as the weak link. This is the sovereign metric layer the generation stack (our nx_qmatvec/nx_qlayer kernels -> VAE -> DiT) is judged by: how close is the rebuilt image to the reference? MSE (exact integer), max-abs-error (the worst pixel -- catches local failures an average hides), and PSNR in dB (the rate-distortion quality measure). Verdict: LOSSLESS (bit- exact) / NEAR-LOSSLESS / LOSSY. Pixels are 0..255 per channel, stored one per slot. license_tier: ORIGINAL Refs: PSNR/MSE (rate-distortion fidelity); SSIM lineage.