open-topoqa-scorer

The scoring half of the open TopoQA vertical, packaged for the same reason as open-topoqa-featurizer-recipe: the upstream implementation is unlicensed, so the only redistributable version is one written from the paper. The architecture came from the Methods section — edge-conditioned multi-head attention updating both nodes and edges, half-width pooled edges, a three-linear readout, sigmoid, MSE loss — and nothing else did. The paper specifies no training or capacity hyperparameters at all, so heads, width, layers, dropout, learning rate, and epochs are ours. See open-topoqa-scorer-environment and topoqa-interface-quality.

The weights are a retrain on a reassembled MAF2 and Dockground corpus. There is no upstream checkpoint to match, and divergence from the released one is the correct outcome rather than a failure: measured against a corrected (x, y, z) TopoQA, the retrain matches or beats every correlation and reaches HAF2 ranking-loss parity.

The recipe declares pytorch and pytorch_geometric explicitly because the ProteinGAT model imports them directly rather than inheriting them. Its open-topoqa-featurizer run dependency is on no public channel yet, so the consuming fixture stages both as sibling path recipes and drags a full torch closure behind them. MIT and noarch, so both could go to a channel together.

Verified green with a solved linux-64 lock.