{
  "dataset": "PathMNIST (9-class colon pathology, 64x64)",
  "num_train": 800,
  "num_val": 200,
  "num_test": 200,
  "epochs": 25,
  "hidden_dim": 64,
  "k_nn": 6,
  "patch_size": 8,
  "patch_stride": 6,
  "nodes_per_graph": 100,
  "feature_dim": 38,
  "main_results": {
    "TopoHyper": {
      "test_acc": 0.820,
      "test_f1_macro": 0.7116,
      "val_acc": 0.830,
      "params": 94858
    },
    "GCN": {
      "test_acc": 0.815,
      "test_f1_macro": 0.7096,
      "val_acc": 0.830,
      "params": 17161
    },
    "Simplicial": {
      "test_acc": 0.815,
      "test_f1_macro": 0.7096,
      "val_acc": 0.810,
      "params": 17417
    },
    "HGNN": {
      "test_acc": 0.810,
      "test_f1_macro": 0.6848,
      "val_acc": 0.795,
      "params": 17161
    },
    "SimpleHybrid": {
      "test_acc": 0.785,
      "test_f1_macro": 0.6731,
      "val_acc": 0.785,
      "params": 14537
    },
    "GAT": {
      "test_acc": 0.770,
      "test_f1_macro": 0.6765,
      "val_acc": 0.780,
      "params": 17801
    }
  },
  "ablation_results": {
    "Full (Bridge + Attention)": {
      "test_acc": 0.820,
      "val_acc": 0.830
    },
    "No Attention (Bridge only)": {
      "test_acc": 0.830,
      "val_acc": 0.835
    },
    "Neither (Average fusion)": {
      "test_acc": 0.805,
      "val_acc": 0.810
    },
    "No Bridge (Attention only)": {
      "test_acc": 0.785,
      "val_acc": 0.790
    }
  },
  "key_findings": [
    "TopoHyper achieves best test accuracy (82.0%) among all models",
    "Bridge matrix is the most critical component - removing it drops accuracy by 3.5%",
    "Cross-structure attention provides modest gains (+1.5%) when bridge is present",
    "Simple concatenation hybrid (SimpleHybrid) underperforms standalone baselines - naive fusion hurts",
    "GAT is the weakest baseline, suggesting attention alone is insufficient for this task",
    "The bridge-only variant (83.0%) actually outperforms the full model - simpler may be better"
  ]
}