diff --git a/ccflow/utils/tokenize.py b/ccflow/utils/tokenize.py
index 91444488..26e8f700 100644
--- a/ccflow/utils/tokenize.py
+++ b/ccflow/utils/tokenize.py
@@ -159,6 +159,13 @@ def _normalize_date(obj):
 
 @normalize_token.register(datetime)
 def _normalize_datetime(obj):
+    cls = type(obj)
+    # pandas.Timestamp subclasses datetime and must not share a token with an equal plain datetime.
+    # Matching by name keeps `import ccflow` from paying for pandas just to register one handler.
+    # DataFrame/Series/Index have no structural handler and fall through to cloudpickle; adding them
+    # (as dask does) would require real lazy registration rather than this name check.
+    if cls is not datetime and any(c.__name__ == "Timestamp" and _module_in(c.__module__, ("pandas",)) for c in cls.__mro__):
+        return ("pd_timestamp", obj.isoformat())
     return ("datetime", obj.isoformat())
 
 
@@ -371,22 +378,7 @@ def _normalize_np_scalar(obj):
         return ("np_scalar", str(type(obj).__name__), obj.item())
 
 
-def _register_pandas() -> None:
-    try:
-        import pandas as pd
-    except ImportError:  # pragma: no cover
-        return
-
-    # Only Timestamp has a structural handler; DataFrame/Series/Index fall through to the cloudpickle
-    # fallback. That works today but is fragile across pandas version upgrades — a follow-up could add
-    # structural handlers (matching dask) for stability and a perf win on large frames.
-    @normalize_token.register(pd.Timestamp)
-    def _normalize_pd_timestamp(obj):
-        return ("pd_timestamp", obj.isoformat())
-
-
 _register_numpy()
-_register_pandas()
 
 
 def tokenize(*args: Any, **kwargs: Any) -> str: