Update dataset card with hold-out experiment results

README.md

@@ -190,6 +190,43 @@ All columns from `nodes.csv` plus:
 - **DAG layers**: Kahn's algorithm; cycle nodes get layer=-1; namespace graph condensed via SCC
 - **Namespace cycles**: 6,055 of 10,097 namespaces in 38 SCCs (largest: 5,899 nodes)
 
+## Hold-Out Experiments
+
+We validate the premise retrieval results with two levels of hold-out experiments to assess information leakage. In the original experiment, all network features (degree, PageRank, betweenness, community, DAG layer) are precomputed on the full graph. Hold-out experiments recompute features on a reduced graph to test whether full-graph computation inflates AUC.
+
+### Edge-level hold-out
+- Randomly remove 20% of edges, recompute all features on remaining 80% graph
+- Tests whether full-graph feature computation inflates AUC
+
+### Declaration-level hold-out
+- Split theorems 80/20, remove ALL edges (incoming and outgoing) of test theorems from training graph
+- Test theorems have zero degree in training graph
+- Simulates the real scenario: predicting premises for a brand new theorem
+
+### Results (Split 1, seed=42)
+
+| Method | AUC (original) | AUC (edge hold-out) | AUC (decl hold-out) |
+|--------|---------------|---------------------|---------------------|
+| Random | 0.499 | 0.497 | 0.500 |
+| Same module | 0.563 | 0.562 | 0.563 |
+| Same namespace | 0.590 | 0.590 | 0.592 |
+| Same community | 0.768 | 0.755 | 0.500 |
+| Network features | 0.991 | 0.988 | 0.978 |
+| All features | 0.994 | 0.992 | 0.984 |
+
+The community feature drops to random (0.500) in declaration-level hold-out because test declarations become isolated nodes with unique community IDs. Despite this, the combined network feature model retains AUC=0.978, confirming that degree, PageRank, and DAG position carry genuine predictive signal independent of information leakage.
+
+Full 5-split results with mean and std are in `experiments/holdout_decl_level.json` (updated incrementally).
+
+### Experiment Files
+
+| File | Description |
+|------|-------------|
+| `experiments/premise_retrieval_results.json` | Original experiment: 6 methods, 4 metrics, 95% CI |
+| `experiments/premise_retrieval_hard_negatives.json` | Hard negatives (same-community) variant |
+| `experiments/holdout_edge_level.json` | Edge-level hold-out (1 split) |
+| `experiments/holdout_decl_level.json` | Declaration-level hold-out (5 splits, incremental) |
+
 ## License
 
 Apache 2.0