ts3_relative_original56 / TRAIN20

6000 full 20-case updates; 982884 parameters; best_train update 6000. All cases U and VM <=1%: False. No heldout or generalization claim.

Final6000 metricMean (%)Worst (%)
U_relative_L21.6630223.011811
VM_relative_L23.0426995.119224
peak5_VM_relative_L22.7921448.188646
VM_max_relative_error2.87036516.947315
VM_at_true_peak_relative_error2.57730917.351450

Negative VM cases: 0; minimum 3.362955985727976 MPa. Raw values preserved.

TRAIN20 only. CPU checkpoint/metrics audit; no CPU full-model forward replay, GPU update replay, FE re-solve or heldout generalization claim.

Both prediction sets: per-case CSV | CPU verification

actual_training_curves

Vector PDF

all20_final6000_errors

Vector PDF

worst_case_fields

Vector PDF

true_peak_location_zoom

Vector PDF