It collects every link from the benchmark work in one index, then goes model by model: Orb v3 (fast, trust it on intermetallics, don't trust it alone on ionic conductors), MACE-MP (best symmetry fidelity on the pass cases, crashes on big spinel cells), CHGNet (the only one that held the Li₆PS₅Cl argyrodite, but not immune), ALIGNN and mCGCNN (property classifiers, not potentials — a structure-only model can't determine magnetic ground states), and the spin-aware frontier (mMTP, SpinGNN++, NEP+SPIN), where the key interface fact is that magnetic order is supplied by you, not inferred from a CIF.
The organizing result is that symmetry preservation tracks bonding chemistry, not space group — same F-43m space group, opposite outcome for half-Heuslers vs argyrodite. There's a decision table near the end mapping material class to tool, and a three-layer validation stack (sanity card → energy gate → symmetry check) that exists because several of our own early "failures" turned out to be broken inputs.
If you run any of your generated candidates through these and something breaks in a way the article doesn't cover, that's exactly what the MLIP failure benchmark quest