mgn: 1000-case geometry and pressure experiment

800 train / 100 validation / 100 test. 6000 actual Adam updates (20 equally weighted train cases per update, 150 train passes). Checkpoint 5900 selected only by validation.

Negative predicted VM cases: 2; raw predictions preserved. Budget completion is not an accuracy pass. These are research results for the specified geometry family and linear FE model; production eligibility remains false.

All 100 test original fields, targets and predictions independently checked on CPU. Train/validation summaries aggregate saved all-case evaluations. The serial worker's exported target has a normalization round trip; metrics here use original FP64 FE labels.

Actual training and validation curves

Vector PDF

All independent test errors

Vector PDF

Worst common-case native fields

Vector PDF

Local view at worst VM point

Vector PDF

All 100 test case metrics CSV · Independent CPU verification

Three-rank terminal saved-state audit: original model and Adam strict shape load, all three CPU/NumPy/Python RNG states restored, CUDA RNG structure checked, original frozen sources and exact 5085-update recovery history verified. No local full-model CPU forward or CUDA optimizer-update replay was performed.