nx_imatrix.nx

nx_imatrix.nx -- IMPORTANCE-MATRIX quantization = rate-distortion WATER-FILLING (the deepest survived task from the governed loop: VRAM-research actionable A1 ~= historic actionable HA1). Naive quantization minimizes RAW error sum((w-wq)^2) and spends bits uniformly. But not every weight matters equally to the output -- an importance matrix (from calibration activations) says how much each weight's error costs. The right objective is the IMPORTANCE-WEIGHTED error sum(imp_i * (w_i - wq_i)^2), and the right bit allocation is Shannon WATER-FILLING: give the next bit to whichever group reduces the weighted error most. This is rate-distortion optimization (Radio), and it is why imatrix cuts low-bit error ~31% at ~zero size cost: same total bits, but spent where distortion is expensive. Distortion here is the QUALITY-relevant (importance-weighted) distortion, not raw MSE -- closing one notch toward the perceptually-meaningful objective the Critic insisted on. license_tier: ORIGINAL Refs: ggml imatrix PR#4930; Radio (rate-distortion); Shannon water-filling.