Predict the QM9 HOMO-LUMO gap below chemical accuracy on the standard 110k/10k/10k split

QM9 contains 133885 small organic molecules (up to 9 heavy atoms C/N/O/F) with DFT (B3LYP/6-31G(2df,p)) ground-state geometries and 12+ computed properties. Train a model to predict the HOMO-LUMO gap $\varepsilon_{\text{gap}} = \varepsilon_{\text{LUMO}} - \varepsilon_{\text{HOMO}}$ (in eV) from molecular structure, using the community-standard split: remove the 3054 molecules that failed the geometric consistency check, then use 110000 molecules for training, 10000 for validation, and 10000 for testing. Report the test-set mean absolute error (MAE) of the gap in meV. Chemical accuracy corresponds to $\approx 43$ meV (1 kcal/mol). Can an agent-built model reach a gap MAE below chemical accuracy, and how close to the reported state of the art ($\approx 30$ meV or below) can it get on a single workstation GPU?

Acceptance. FULLY RESOLVES: a model achieving test-set HOMO-LUMO gap MAE $\le 30$ meV on the standard 110000/10000/10000 QM9 split (with the 3054 uncharacterized molecules removed), reported with the exact split indices or a deterministic seed/recipe, model architecture and hyperparameters, and a runnable training+evaluation script and weights. PARTIAL: any reproducible result below chemical accuracy ($\le 43$ meV) with full split disclosure, or a documented negative result for a specified architecture. Metric: MAE of $\varepsilon_{\text{gap}}$ over the 10000-molecule test set, in meV. The split protocol must be stated exactly so the number is comparable.