fullseye

Accuracy benchmark — champion vs null (holdout)

Short-budget sweep: gens=8, pop=10, n_train=6, n_holdout=4, size=32, seed=0. Selection is on TRAIN only; all scores below are on the HELD-OUT set. Higher is better. Not the headline long-budget numbers — this is the standing ‘does evolution beat the null’ table.

problem unit trivial hand random champion best null beats null? gap
denoise dB PSNR 15.2785 20.2116 12.3234 16.3677 20.2116 no 1.1763
edge F1 0.4172 0.9324 0.2851 0.3817 0.9324 no 0.2283
binarize IoU 0.7756 0.8865 0.2626 0.5714 0.8865 no 0.0584
count 1/(1+err) 0.0642 0.625 0.5833 0.625 0.625 no 0.2917
locate 1/(1+px) 0.0 1.0 0.0487 1.0 1.0 no 0.0
locate_rot 1/(1+px) 0.0 1.0 0.185 0.5573 1.0 no -0.1455
classify accuracy 0.5 0.5 1.0 1.0 1.0 no 0.0
barcode 1/(1+err) 0.5 1.0 1.0 1.0 1.0 no 0.0
vol_denoise dB PSNR 19.3031 21.3223 20.3518 24.5895 21.3223 yes -1.029
vol_count 1/(1+err) 0.2986 1.0 1.0 1.0 1.0 no 0.0

Champion beats the best null on holdout in 1/10 problems. A negative gap means holdout ≥ train (no overfit); a large positive gap flags overfitting.