New models delivered: 4/4. Completed TS3 and FNO references are reused. Final6000 metrics appear only after SHA download, independent CPU audit and visual delivery.
Identical TRAIN20 maximin cases, original56 inputs/4 outputs, seed20261029, FP64, relative-L2-squared U/VM loss, 6000 actual20-case updates. Active trainable counts matched within1%. TRAIN20 fitting, not heldout generalization; one seed and fixed budget do not establish universal causality.
| Model | Status / updates | Active parameters | U mean / worst (%) | VM mean / worst (%) | Peak5 VM mean / worst (%) | Negative VM cases |
|---|---|---|---|---|---|---|
| TS1 / Transolver | delivered / 6000 | 982884 | 1.5197/3.0301 | 3.0522/5.0910 | 2.7094/7.9710 | 0 |
| TS2 / Transolver++ | delivered / 6000 | 981172 | 1.5800/3.3347 | 3.2820/5.2455 | 2.8295/8.0232 | 0 |
| TS3 reference | delivered / 6000 | 982884 | 1.6630/3.0118 | 3.0427/5.1192 | 2.7921/8.1886 | 0 |
| GNOT | delivered / 6000 | 982170 | 2.1814/3.0846 | 7.2196/8.8306 | 8.0863/9.9366 | 2 |
| OFormer | delivered / 6000 | 990268 | 1.6468/1.9883 | 3.4072/4.5346 | 2.5496/3.2298 | 0 |
| FNO reference | delivered / 6000 | 984668 | 2.0247/2.9482 | 1.8390/2.3637 | 1.0444/1.9497 | 2 |
Raw negative predictions are retained. None of the currently delivered configurations satisfies U and VM <=1% on every training case. Best-train checkpoints are supplementary.
Full metrics CSV | Delivery evidence
At the fixed6000 budget, TS1 / Transolver has the lowest mean U error (1.5197%), while OFormer has the lowest worst-case U error (1.9883%). These are different criteria; there is no single winner on all metrics.
FNO reference has the lowest mean VM error (1.8390%) and FNO reference the lowest worst-case VM error (2.3637%). FNO nevertheless has 2 cases with raw negative VM, so its lower relative error does not guarantee physically nonnegative predictions.
Compared with the TS3 reference, TS1, TS2 and OFormer have lower mean U errors but higher mean VM errors. GNOT has higher mean U and VM errors than TS3 under this exact setup. The official architectures with matched parameter counts still behave differently at the same fixed budget; this observation does not establish a universal architectural cause.
GNOT final6000: mean/worst U 2.1814%/3.0846%, mean/worst VM 7.2196%/8.8306%, 2 negative-VM cases, minimum -4.703556 MPa. The stress field remains the main error component. The completed budget was not extended to reach an accuracy threshold.
No configuration meets both U and VM <=1% on every TRAIN20 case. These 20 cases are training cases, not an independent test set. One seed, one parameter scale and one fixed optimization budget support this comparison only; the separate 800/100/100 GNOT experiment remains pending and will have its own heldout-test report.