dataset N=12000 target_scale=14.595 cond=(a,b,logρ) dev=cuda step 0 loss 0.9989 step 2000 loss 0.0117 step 4000 loss 0.0177 step 6000 loss 0.0146 step 8000 loss 0.0111 step 10000 loss 0.0058 trained in 8.3s final loss 0.0113 === reference plant G=1/(s+1): cost-CONDITIONED single-shot generation vs analytic === rho=0.1 single-shot +0.43% best-of-16 +0.21% RMSE(φ-φ*)=0.240 rho=0.3 single-shot +0.43% best-of-16 +0.18% RMSE(φ-φ*)=0.168 rho=1 single-shot +1.75% best-of-16 +0.26% RMSE(φ-φ*)=0.175 rho=3 single-shot +3.12% best-of-16 +1.29% RMSE(φ-φ*)=0.157 rho=10 single-shot +5.09% best-of-16 +1.62% RMSE(φ-φ*)=0.150 held-out plants (a,b~U[0.5,2]) single-shot median regret @rho=1: +1.48% saved -> results/siso_coherent/costcond_siso.json, models/costcond_siso.pth