Adaptive MSE
–
Uniform-16 MSE (baseline)
–
MSE reduction
–
Win rate (rows adaptive wins)
–
MSE per kurtosis bin (green = adaptive, grey = uniform)
How it works
1. Stats: per row compute excess kurtosis mean(z^4)-3, skew, outlier fraction (|z|>3). 2. Switch rule: kurtosis > target (slider) or outlier-frac > 0.02 → odd16 companded grid (fine near zero); |skew| > 1 → shifted grid; else uniform16. 3. Scale: closed-form LS s = (x·c)/(c·c) with assign→rescale→reassign→rescale. Mirrors kurt_switch/core.py: compute_row_stats, select_codebook, quantize_row.
Code folder: github.com/DastanZar/kurt-switch · portfolio card