Citation: drop venue

Update model card for the revised paper

Publish only the checkpoints the paper's tables need

Model card: lead with the UMA physical reward (93.01/93.48 across seeds at NFE 32, 94.05 at 50), add the full NFE grid, the corrector chain on the UMA weights, the scale-up negative result, and a warning against cross-protocol comparison

CrystAF after the scaled-up UMA-reward RL (240 families/epoch, group 150), epoch 4 -- the peak of that run, PB 93.29 at NFE=32; a negative result, it matches the 80-family recipe after the same number of updates

MolCrystalFlow after UMA-guided NFT (lattice-volume RMAD 3.88 -> 3.16)

Clari-L after UMA-guided NFT (PB 86.88 -> 91.37 at 1000 families)

Clari-M after UMA-guided NFT (PB 88.43 -> 91.61 at 1000 families)

CrystAF after UMA-reward RL rolling out on the NFE=32 grid, epoch 12 (best packing: vol 1.81, PDD 10.14)

CrystAF after UMA-reward RL, seed 929 epoch 6 (PB 93.01, clash 7.98)

CrystAF after UMA-reward RL, seed 2029 epoch 6 (PB 93.48 at NFE=32)

Update model card: two result tables (with/without stereo correction), Stage-2 weights, reproducibility caveat

Second seed of the same recipe (reproducibility check)

CrystAF-NFT (MeanFlowNFT mfpure) epoch 12 report weights

Upload README.md with huggingface_hub

Upload teacher-rank800-pbnft-epoch1.pt with huggingface_hub

Upload crystaf-cont3-step2000.pt with huggingface_hub

initial commit