| Parameter | Primary path | Secondary |
d_model | MSA -> Embed -> Evo | broad downstream |
n_heads | Evo -> Struct | strong |
n_layers | Evo full stack | moderate struct |
recycling | recycle loop | Struct -> Output |
lr | Evo + training | wide, nonlinear |
Geometry Contract
Implemented: Nussinov DP (O(n³), WC+GU), Frenet-style backbone evolution, Vietoris-Rips filtration sweep, 4-type graph edge encoding.
Interaction invariant: sequence selection is synchronized across secondary/3D/TDA/action views.
Next wire: external 3D predictor as coordinate source and direct VR fingerprint refresh.
coords = protenix_pred(seq, use_rna_msa=True, seeds=[101,102,103])
h0,h1,betti = vr_persistence(coords, eps_max=40, steps=60)
node_feats = [one_hot(nt), confidence[i], coords[i]]