Cocktail sweep (71/72 combos, jobs torch:14797484 + torch:14830472): interaction_context_dc_penalty {0, 1e-4, 1e-3, 1e-2} x interaction_feature_l2 {1e-4, 1e-2} x filter_transform capacity {1x32, 2x64, 2x128} x seeds {0..3}, 101 epochs, order-2 onset at epoch 50, on PURELY FIRST-ORDER ten-motif no-noise data (GT order-2 mass exactly 0, so all model order-2 mass is spurious). FINDING 1 -- CAPACITY IS A SEED-ROBUSTNESS LEVER, NOT A FIT… See the full description on the dataset page:
https://huggingface.co/datasets/arushram/fnbm-current-cocktail-dc-capacity-20260726.
