diff --git a/test/io/test_esmf.py b/test/io/test_esmf.py
index 552bf0009..87d9b22d0 100644
--- a/test/io/test_esmf.py
+++ b/test/io/test_esmf.py
@@ -35,6 +35,68 @@ def test_read_esmf_dataset(gridpath, datasetpath):
     for dim in dims:
         assert dim in uxds.dims
 
+@pytest.mark.parametrize("mask_and_scale", [True, False])
+@pytest.mark.parametrize(
+    "num_element_conn", [[4, 3, 3], [4, 4, 4]], ids=["counted", "miscounted"]
+)
+def test_read_esmf_padding_independent_of_cf_decoding(
+    num_element_conn, mask_and_scale, tmp_path
+):
+    """Padding is recognized whether or not xarray decoded the fill value.
+
+    ESMF pads a short face in `elementConn` with that variable's `_FillValue`.
+    With CF decoding on, xarray replaces it with NaN and promotes the array to
+    float; with decoding off, the raw -1 comes through. Casting first and checking
+    for INT_FILL_VALUE afterwards recognizes neither: the raw -1 becomes the index
+    -2, and the NaN cast is platform-dependent -- arm64 gives 0, so the padding
+    decodes to -1, a negative index that silently wraps to the last node.
+
+    So the padding has to be located before the cast, from `numElementConn`.
+    """
+    node_lon = np.array([0.0, 120.0, 120.0, 0.0, 240.0])
+    node_lat = np.array([0.0, 0.0, 10.0, 10.0, 5.0])
+
+    # 1-based and -1 padded, as ESMF specifies: one quad and two triangles
+    in_ds = xr.Dataset(
+        {
+            "nodeCoords": xr.DataArray(
+                np.column_stack([node_lon, node_lat]),
+                dims=("nodeCount", "coordDim"),
+                attrs={"units": "degrees"},
+            ),
+            "elementConn": xr.DataArray(
+                np.array([[1, 2, 3, 4], [2, 5, 3, -1], [1, 4, 5, -1]], dtype=np.int32),
+                dims=("elementCount", "maxNodePElement"),
+                attrs={"_FillValue": np.int32(-1)},
+            ),
+            "numElementConn": xr.DataArray(
+                np.array(num_element_conn, dtype=np.byte), dims="elementCount"
+            ),
+        }
+    )
+
+    path = tmp_path / "esmf_ragged.nc"
+    in_ds.to_netcdf(path)
+
+    with xr.open_dataset(path, mask_and_scale=mask_and_scale) as raw:
+        uxgrid = ux.open_grid(raw)
+
+    np.testing.assert_array_equal(
+        uxgrid.face_node_connectivity.values,
+        np.array([
+            [0, 1, 2, 3],
+            [1, 4, 2, INT_FILL_VALUE],
+            [0, 3, 4, INT_FILL_VALUE],
+        ]),
+    )
+    # The connectivity knows each face's real size whatever the file claimed
+    real_sizes = (uxgrid.face_node_connectivity.values != INT_FILL_VALUE).sum(axis=1)
+    np.testing.assert_array_equal(real_sizes, [4, 3, 3])
+
+    # "n_nodes_per_face" is still the declared count, so a miscount survives here
+    np.testing.assert_array_equal(uxgrid.n_nodes_per_face.values, num_element_conn)
+    assert uxgrid.n_node == 5
+
 def test_esmf_round_trip_consistency(gridpath):
     """Test round-trip serialization of grid objects through ESMF xarray format.
 
diff --git a/uxarray/io/_esmf.py b/uxarray/io/_esmf.py
index 59ad7ea3c..0ec4e3f16 100644
--- a/uxarray/io/_esmf.py
+++ b/uxarray/io/_esmf.py
@@ -93,12 +93,30 @@ def _read_esmf(in_ds):
         # assume start index is 1 if one is not provided
         start_index = 1
 
-    face_node_connectivity = in_ds["elementConn"].astype(INT_DTYPE)
-    face_node_connectivity = xr.where(
-        face_node_connectivity != INT_FILL_VALUE,
-        face_node_connectivity - start_index,
-        face_node_connectivity,
+    element_conn = in_ds["elementConn"]
+    face_dim, node_dim = element_conn.dims
+
+    # "numElementConn" gives the face size, so locate the padding positionally.
+    # Matching the sentinel means guessing: CF decoding turns it into NaN, and the
+    # cast below preserves neither NaN nor the raw value as INT_FILL_VALUE.
+    positions = xr.DataArray(
+        np.arange(element_conn.sizes[node_dim], dtype=INT_DTYPE), dims=node_dim
     )
+    fill_mask = (positions >= n_nodes_per_face).transpose(face_dim, node_dim)
+
+    # NaN is never a usable index, whatever "numElementConn" claims
+    # NaN is never a usable index, whatever "numElementConn" claims
+    fill_mask = fill_mask | element_conn.isnull()
+
+    # ...and neither is the declared sentinel, whatever "numElementConn" claims
+    sentinel = element_conn.attrs.get(
+        "_FillValue", element_conn.encoding.get("_FillValue")
+    )
+    if sentinel is not None:
+        fill_mask = fill_mask | (element_conn == sentinel)
+
+    face_node_connectivity = element_conn.fillna(0).astype(INT_DTYPE) - start_index
+    face_node_connectivity = xr.where(fill_mask, INT_FILL_VALUE, face_node_connectivity)
 
     out_ds["face_node_connectivity"] = xr.DataArray(
         data=face_node_connectivity,