diff --git a/tests/test_graph_dict.py b/tests/test_graph_dict.py
new file mode 100644
index 0000000..6454f03
--- /dev/null
+++ b/tests/test_graph_dict.py
@@ -0,0 +1,53 @@
+import pytest
+import sys
+import pathlib
+import importlib.util
+import glob
+import os
+import torch
+
+# allow imports from repo root
+sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[1]))
+
+from compare_encoders import optimizer_to_data
+
+
+def canonical_representation(opt):
+    data = optimizer_to_data(opt)
+
+    def simplify(attrs):
+        result = {}
+        for k, v in attrs.items():
+            key = str(getattr(k, "name", k))
+            result[key] = v.tolist() if hasattr(v, "tolist") else v
+        return result
+
+    nodes = sorted(
+        (int(t), tuple(sorted((k, str(v)) for k, v in simplify(attrs).items())))
+        for t, attrs in zip(data.node_types.tolist(), data.node_attributes)
+    )
+    edges = set(map(tuple, data.edge_index.numpy().T))
+    return nodes, edges
+
+
+@pytest.mark.parametrize(
+    "pt_path", glob.glob(os.path.join("computation_graphs", "optimizers", "*.pt")))
+def test_graph_dict_matches_source(pt_path):
+    base = os.path.splitext(pt_path)[0]
+    py_path = base + ".py"
+    spec = importlib.util.spec_from_file_location("mod", py_path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+
+    cls_candidates = [v for v in module.__dict__.values() if isinstance(v, type) and issubclass(v, torch.nn.Module)]
+    assert len(cls_candidates) == 1, f"Could not find optimizer class in {py_path}"
+    cls = cls_candidates[0]
+
+    scripted_py = torch.jit.script(cls())
+    scripted_pt = torch.jit.load(pt_path)
+
+    nodes_py, edges_py = canonical_representation(scripted_py)
+    nodes_pt, edges_pt = canonical_representation(scripted_pt)
+
+    assert nodes_py == nodes_pt
+    assert edges_py == edges_pt