diff --git a/flyem_snapshot/inputs/rois.py b/flyem_snapshot/inputs/rois.py
index 3539d54..ffb604e 100644
--- a/flyem_snapshot/inputs/rois.py
+++ b/flyem_snapshot/inputs/rois.py
@@ -206,6 +206,14 @@ def load_point_rois(cfg, point_df, roiset_names):
         assert isinstance(roi_ids, dict)
 
         roi_ids = _apply_roi_renames(point_df, roiset_name, roi_ids, roiset_cfg['rename-rois'])
+
+        if '<unspecified>' in point_df[roiset_name].dtype.categories:
+            assert (u := roi_ids.get('<unspecified>')) == 0, \
+                f"Non-zero '<unspecified>' label: {u}"
+        else:
+            assert '<unspecified>' not in roi_ids, \
+                "Did not expect to see '<unspecified>' in roi_ids, since it is not in the categories."
+
         roisets[roiset_name] = roi_ids
 
     _check_duplicate_rois(roisets)
@@ -297,7 +305,7 @@ def _load_columns_for_roiset(roiset_name, roiset_cfg, point_df, dvid_cfg, proces
     if bool(roiset_cfg['labelmap']) != (roiset_cfg['source'] == 'dvid-labelmap'):
         raise RuntimeError("Please supply a labelmap for your ROIs IFF you selected source: dvid-labelmap")
 
-    roi_vol, roi_box, roi_ids = _load_roi_vol(roiset_name, roi_ids, roiset_cfg['labelmap'], dvid_cfg, processes)
+    roi_vol, roi_box, roi_ids = _load_roi_vol_from_cache_or_dvid(roiset_name, roi_ids, roiset_cfg['labelmap'], dvid_cfg, processes)
     extract_labels_from_volume(point_df, roi_vol, roi_box, 5, roi_ids, roiset_name, skip_index_check=False)
 
     return roi_ids
@@ -326,7 +334,6 @@ def _load_roi_unions(roiset_name, src_roiset_name, src_lists, point_df):
 
     Returns:
         dict {roi_name: roi_label_int} for the new roiset.
-        By convention, "<unspecified>" (label 0) is not included in dict.
 
     Note:
         The sub-rois must all come from the same source roiset, and it
@@ -367,14 +374,16 @@ def _load_roi_unions(roiset_name, src_roiset_name, src_lists, point_df):
     # Apply mapping to obtain the new ROIs (union of subrois)
     point_df[roiset_name] = point_df[src_roiset_name].map(union_mapping).fillna('<unspecified>')
     assert point_df[roiset_name].dtype == 'category'
+    point_df[roiset_name] = point_df[roiset_name].cat.remove_unused_categories()
 
     # Produce integers from names to create the _label column
-    roi_ids = {roi: i for i, roi in enumerate(union_mapping.dtype.categories)}
+    union_categories = point_df[roiset_name].dtype.categories
+    start = 0 if '<unspecified>' in union_categories else 1
+    roi_ids = {roi: i for i, roi in enumerate(union_categories, start)}
     label_dtype = np.min_scalar_type(max(roi_ids.values()))
     point_df[f'{roiset_name}_label'] = point_df[roiset_name].map(roi_ids).astype(label_dtype)
 
-    assert roi_ids['<unspecified>'] == 0
-    del roi_ids['<unspecified>']
+    assert ('<unspecified>' not in roi_ids) or (roi_ids['<unspecified>'] == 0)
     return roi_ids
 
 
@@ -388,6 +397,25 @@ def _load_roi_col(roiset_name, roi_ids, point_df):
         - primary_roi_label (integers)
 
     If it contains only one or the other, then we calculate the missing one using roi_ids.
+
+    Args:
+        roiset_name:
+            string
+
+        roi_ids:
+            Optional.
+            If given, must be dict of {name: int}.
+            The {'<unspecified>': 0} entry may be omitted, in which case we'll insert it if necessary.
+            No other ROI is permitted to have label 0.
+            If roi_ids is None, we'll create it.
+
+        point_df:
+            DataFrame, modified IN-PLACE.
+
+    Returns:
+        roi_ids
+            - Possibly created from scratch
+            - Possibly updated with '<unspecified>': 0
     """
     if not roi_ids and roiset_name not in point_df:
         raise RuntimeError(
@@ -406,15 +434,32 @@ def _load_roi_col(roiset_name, roi_ids, point_df):
             eval(f'f"{roi_ids}"', None, {'x': x}): x  # pylint: disable=eval-used
             for x in unique_ids if x != 0
         }
+        if 0 in unique_ids:
+            roi_ids['<unspecified>'] = 0
 
     if roiset_name in point_df.columns:
         point_df[roiset_name] = point_df[roiset_name].fillna("<unspecified>")
-        empirical_names = set(point_df[roiset_name].unique())
+
+        # If it's present, <unspecified> is always listed first.
+        empirical_names = sorted(point_df[roiset_name].unique())
+        if '<unspecified>' in empirical_names:
+            empirical_names.remove('<unspecified>')
+            empirical_names = ['<unspecified>', *empirical_names]
+
         if not roi_ids:
-            roi_ids = {n: i for i, n in enumerate(sorted(empirical_names), start=1)}
-        if '<unspecified>' in empirical_names and '<unspecified>' not in roi_ids:
-            roi_ids['<unspecified>'] = 1 + max(roi_ids.values())
-        if unlisted_rois := empirical_names - set(roi_ids.keys()):
+            start = 0 if '<unspecified>' in empirical_names else 1
+            roi_ids = {n: i for i, n in enumerate(empirical_names, start)}
+
+        if (u := roi_ids.get('<unspecified>', 0)) != 0:
+            raise RuntimeError(f"roiset {roiset_name}: <unspecified> must have label 0, not {u}")
+
+        if '<unspecified>' in empirical_names:
+            roi_ids['<unspecified>'] = 0
+
+        if zero_keys := {k for k,v in roi_ids.items() if v == 0 and k != "<unspecified>"}:
+            raise RuntimeError(f"roiset '{roiset_name}' uses ID 0 for ROIs other than '<unspecified>': {zero_keys}")
+
+        if unlisted_rois := set(empirical_names) - set(roi_ids.keys()):
             raise RuntimeError(
                 f"Your config for ROI column '{roiset_name}' explicitly lists ROI names, but that list is incomplete.\n"
                 f"The following ROIs were found in the data but not in the config:\n"
@@ -438,6 +483,8 @@ def _load_roi_col(roiset_name, roi_ids, point_df):
     if roiset_name in point_df.columns and not isinstance(point_df[roiset_name].dtype, pd.CategoricalDtype):
         point_df[roiset_name] = point_df[roiset_name].astype('category')
 
+    assert roi_ids.get('<unspecified>') == 0 or ('<unspecified>' not in point_df[roiset_name].dtype.categories)
+
     if expected_cols <= {*point_df.columns}:
         # Necessary columns are already present
         return roi_ids
@@ -457,7 +504,7 @@ def _load_roi_col(roiset_name, roi_ids, point_df):
     return roi_ids
 
 
-def _load_roi_vol(roiset_name, roi_ids, roi_labelmap_name, dvid_cfg, processes):
+def _load_roi_vol_from_cache_or_dvid(roiset_name, roi_ids, roi_labelmap_name, dvid_cfg, processes):
     """
     Load an ROI volume, either from a list of DVID 'roi' instances,
     or from a single low-res DVID labelmap instance.
@@ -495,6 +542,8 @@ def _load_roi_vol(roiset_name, roi_ids, roi_labelmap_name, dvid_cfg, processes):
             eval(f'f"{roi_ids}"', None, {'x': x}): x  # pylint: disable=eval-used
             for x in unique_ids if x != 0
         }
+        if 0 in unique_ids:
+            roi_ids['<unspecified>'] = 0
 
     return roi_vol, roi_box, roi_ids
 
@@ -547,14 +596,18 @@ def _apply_roi_renames(point_df, roiset_name, roi_ids, renames):
 
 
 def _check_duplicate_rois(roisets):
-    all_rois = pd.Series([
+    """
+    Raise an error if the same ROI name is present
+    in more than one roiset (except '<unspecified>')
+    """
+    all_rois = [
         k
         for d in roisets.values()
         for k in d.keys()
         if k != '<unspecified>'
-    ])
+    ]
 
-    vc = all_rois.value_counts()
+    vc = pd.Series(all_rois).value_counts()
     duplicate_rois = vc[vc > 1].index.tolist()
     if duplicate_rois:
         raise RuntimeError(f"ROIs duplicated in multiple roisets: {duplicate_rois}")
diff --git a/flyem_snapshot/outputs/neuprint/neuprint.py b/flyem_snapshot/outputs/neuprint/neuprint.py
index 0ffe13b..b6c4a95 100644
--- a/flyem_snapshot/outputs/neuprint/neuprint.py
+++ b/flyem_snapshot/outputs/neuprint/neuprint.py
@@ -311,6 +311,15 @@ def export_neuprint(cfg, point_df, partner_df, element_tables, ann, body_sizes,
     point_df = point_df.loc[point_df['body'] != 0]
     partner_df = partner_df.loc[(partner_df['body_pre'] != 0) & (partner_df['body_post'] != 0)]
 
+    # We don't store neuprint properties for the "<unspecified>" ROI.
+    syn_roisets = copy.deepcopy(syn_roisets)
+    for roi_ids in syn_roisets.values():
+        roi_ids.pop('<unspecified>', None)
+
+    element_roisets = copy.deepcopy(element_roisets)
+    for roi_ids in element_roisets.values():
+        roi_ids.pop('<unspecified>', None)
+
     neuprint_ann = neuprint_segment_annotations(cfg, ann)
 
     point_df, partner_df = restrict_synapses_for_setting(
@@ -347,7 +356,7 @@ def export_neuprint(cfg, point_df, partner_df, element_tables, ann, body_sizes,
 
     connectome = export_neuprint_segment_connections(cfg, partner_df)
 
-    # FIXME: It would be good to verify that there are no duplicated Element points (including Synapses)
+    # TODO: It would be good to verify that there are no duplicated Element IDs (including Synapses)
     export_neuprint_elementsets(cfg, element_tables, connectome)
     export_neuprint_elements(cfg, element_tables, element_roisets)
     export_neuprint_elements_closeto(element_tables)