All labels are provisional CSV-seeded areas; registration is unvalidated. Same-run fit is separate from leave-one-run-out prediction. Three production runs cannot establish production accuracy.
Mesh GNN uses the previous message processor. Residual Geo GNN adds short residual edge messages and position frequencies. Geo MLP has the same position and process input without graph messages. The experiment covers 1 graph budgets and 1 model families, using per-run balanced weak loss.
Values below show residual_geo average precision. Individual and joint values are evaluated on runs used for fitting. Held-out values evaluate an unseen run. The last column trains on the other AC28 run only, allowing a direct comparison with adding ME1a for AC28 holdouts.
| Requested nodes | Evaluated run | Individual fit AP | Joint fit AP | Held-out AP | AC28-only held-out AP |
|---|---|---|---|---|---|
| 5,000 | AC28_16h | 0.3298 | 0.1757 | 0.1020 | 0.1772 |
| 5,000 | AC28_18h01 | 0.0380 | 0.0307 | 0.0133 | 0.0203 |
| 5,000 | ME1a_14h06 | 0.0685 | 0.0920 | 0.0010 | — |
| Requested nodes | Model | Training | Evaluated run | Mapped AP | Mapped IoU | Weak-label transfer IoU |
|---|---|---|---|---|---|---|
| 5,000 | residual_geo | joint | AC28_16h | 0.1757 | 0.1765 | 0.2667 |
| 5,000 | residual_geo | joint | AC28_18h01 | 0.0307 | 0.0455 | 0.0345 |
| 5,000 | residual_geo | joint | ME1a_14h06 | 0.0920 | 0.0506 | 0.1750 |
| 5,000 | residual_geo | holdout_AC28_16h | AC28_16h | 0.1020 | 0.1364 | 0.2667 |
| 5,000 | residual_geo | holdout_AC28_18h01 | AC28_18h01 | 0.0133 | 0.0286 | 0.0345 |
| 5,000 | residual_geo | holdout_ME1a_14h06 | ME1a_14h06 | 0.0010 | 0.0000 | 0.1750 |
Full fit and per-run metrics CSV · Metrics and checkpoints JSON · Main AC28 + ME1a report