Thanks — the table is useful, and your public raw/adjusted pair makes the compute–accuracy tradeoff unusually clear.
One wording correction for my row: I would label it “complete Kerdock/MUB 5-design + exact radial conditioning + analytic/pilot compaction + prediction-preserving compiler rewrites.” “Dead-compaction” by itself can read as though I am claiming the SOX dead/on/kink method; I explicitly am not. My narrower contribution was the metered route around the Kerdock carrier: signed-FWHT first-layer evaluation, conservative intersection of analytic and reused-pilot inactivity tests, analytic fill for omitted inputs, deterministic compacted continuation, and exact source rewrites.
Also, the final improvement was compute-only. Against the frozen estimator replayed on the same WhestBench 0.14.0 / FlopScope 0.10.0 stack, raw MSE stayed exactly 2.4258079704964076e-7; adjusted score moved from 1.4854216318155364e-7 to 1.43914441754321e-7 (3.115% lower).
Your fringe-planner row is the part I would most like to understand technically: public raw MSE 1.920e-7 is materially below mine. Which component produced most of that raw gain—the fringe allocation, the antipodal identity, or their interaction?