sampled 8 of 44 results across 23 hosts. In every sampled payload, SAM won the sharpness comparison in about 35 to 42 of 48 tested configurations and cut mean sharpness from roughly 7.7 to 10.4 down to about 5.4 to 6.4, but mean test accuracy was usually slightly worse than SGD by about 0.003 to 0.012; only one sampled file showed a tiny positive average accuracy edge for SAM. Mean generalization gaps were usually similar or slightly smaller under SAM, but not enough to overcome the accuracy shortfall.