diff --git a/smauglab/transforms/synthseg/functional.py b/smauglab/transforms/synthseg/functional.py
index 93d9cc9..954f8f7 100644
--- a/smauglab/transforms/synthseg/functional.py
+++ b/smauglab/transforms/synthseg/functional.py
@@ -670,15 +670,28 @@ def distinct_values(volume: torch.Tensor) -> torch.Tensor:
 
 
 def _label_counts_per_sample(label_map: torch.Tensor) -> torch.Tensor:
-    """`(B, max_label + 1)` voxel counts per label, per sample."""
+    """`(B, max_label + 1)` voxel counts per label, per sample. Negative values are not counted.
+
+    A negative value is not a label. nnU-Net writes -1 for voxels outside the annotated
+    region, and `clamp_min(0)` counted every one of them as background -- while
+    `em_subdivide_labels`, the only caller, then selects that label's voxels with
+    `label_map == p`. The count exceeded the number of voxels selected, so the
+    `randint(0, count)` subsample indexed past the end of the gathered values. On CUDA
+    that is an asynchronous `index out of bounds` device-side assert, reported at
+    whichever kernel happens to run next, and it only fires when an oversized draw is
+    actually taken -- so it presented as an occasional, unattributable crash.
+    """
     batch = label_map.shape[0]
-    flat = label_map.reshape(batch, -1).clamp_min(0)
-    n_bins = int(flat.max()) + 1
+    flat = label_map.reshape(batch, -1)
+    keep = flat >= 0
+    # The clamp is on the offset only, so a negative label cannot reach another sample's
+    # bins; `keep` is what removes it from the histogram.
+    n_bins = int(flat.max().clamp_min(0)) + 1
     # One `bincount` over row-offset labels: a `scatter_add_` into a handful of bins
     # has every thread fighting for the same addresses, where bincount accumulates
     # in shared memory per block.
-    offset = flat + torch.arange(batch, device=flat.device).view(-1, 1) * n_bins
-    return torch.bincount(offset.reshape(-1), minlength=batch * n_bins).view(batch, n_bins)
+    offset = flat.clamp_min(0) + torch.arange(batch, device=flat.device).view(-1, 1) * n_bins
+    return torch.bincount(offset[keep], minlength=batch * n_bins).view(batch, n_bins)
 
 
 def _em_gmm_1d(x_fit: torch.Tensor, n_components: int, n_iters: int, eps: float) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None:
@@ -805,6 +818,13 @@ def em_subdivide_labels(
     label_counts = _label_counts_per_sample(label_map)
     totals = label_counts.sum(dim=0)
     parents = totals.nonzero().flatten().tolist()  # sorted, batch-wide
+    if not parents:
+        # No non-negative label anywhere: every voxel is outside the annotated region.
+        # One undivided region, no EM, and a merge map that resolves -- rather than
+        # indexing an empty `parents` in the return statement at the end of this
+        # function. Previously unreachable, because the histogram clamped the negative
+        # labels into the background bin.
+        return torch.zeros_like(label_map), [0], [int(background_label)]
     counts_per_sample = label_counts.tolist()
     parent_to_idx = {p: i for i, p in enumerate(parents)}
     lo, hi = int(background_clusters_range[0]), int(background_clusters_range[1])
@@ -814,7 +834,7 @@ def em_subdivide_labels(
     # decoding shifted every label by +1, so the real background is no longer 0), fall back
     # to the largest-area label, so it still receives the richer [min, max] background
     # split rather than the 2-cluster foreground split.
-    if background_label not in parents and len(parents) > 0:
+    if background_label not in parents:
         background_label = int(totals.argmax())
 
     bg_n_shared = int(torch.randint(lo, hi + 1, (1,), device=device)) if same_on_batch else None
diff --git a/unit_tests/test_synthseg_em_ignore_label.py b/unit_tests/test_synthseg_em_ignore_label.py
new file mode 100644
index 0000000..e09cbb5
--- /dev/null
+++ b/unit_tests/test_synthseg_em_ignore_label.py
@@ -0,0 +1,102 @@
+"""EM label completion on a label map that carries nnU-Net's ignore label.
+
+nnU-Net writes -1 for voxels outside the annotated region, and the preprocessed
+volumes of a partially-annotated dataset are full of them. `em_subdivide_labels`
+took each label's voxel count from a histogram of `label_map.clamp_min(0)`, which
+counts every -1 as background, and then selected that label's voxels with
+`label_map == p`, which does not. The count was larger than the selection, so the
+`randint(0, count)` subsample used to fit the background GMM could draw an index
+past the end of the gathered values.
+
+On CPU that is an `IndexError`. On CUDA it is an asynchronous `index out of bounds`
+device-side assert, attributed to whichever kernel happens to run next, and it
+only fires when an oversized index is actually drawn -- for a 128^3 patch with six
+ignore voxels and 100 000 fit draws, about two runs in five. The symptom was an
+occasional crash that pointed at innocent code.
+
+The first test is the invariant itself and is deterministic; the rest exercise the
+call that used to raise, and the all-ignore patch that the clamp made
+accidentally representable.
+"""
+
+from __future__ import annotations
+
+import torch
+
+from smauglab.transforms.synthseg import functional as FN
+from unit_tests.helpers import SmaugLabTestCase
+
+SHAPE = (16, 16, 16)
+VOXELS = SHAPE[0] * SHAPE[1] * SHAPE[2]
+FOREGROUND = 8 * 8 * 8
+#: Smaller than the number of background voxels, so the subsample path is the one
+#: under test: with `max_fit_voxels >= count` the function uses every voxel and
+#: never draws an index at all.
+MAX_FIT = 64
+#: Enough ignored voxels that an out-of-range draw is a near-certainty rather than a
+#: coin flip: the old histogram would report 3584 background voxels where only 1584
+#: can be selected, so P(no bad index in 64 draws) = (1584/3584)^64 < 1e-23.
+N_IGNORE = 2000
+
+
+def pair_with_ignore(n_ignore: int = N_IGNORE, batch: int = 2):
+    """A label map with background, one structure and `n_ignore` ignored voxels."""
+    depth, height, width = SHAPE
+    labels = torch.zeros(batch, 1, depth, height, width, dtype=torch.long)
+    labels[..., 4:12, 4:12, 4:12] = 1
+    labels.reshape(batch, -1)[:, :n_ignore] = -1
+    image = torch.rand(batch, 1, depth, height, width)
+    return labels, image
+
+
+class EmIgnoreLabelTests(SmaugLabTestCase):
+    def test_counts_match_the_selection_they_are_used_for(self):
+        """The histogram must count exactly the voxels `label_map == p` selects.
+
+        This is the invariant the crash violated, checked directly so a future
+        rewrite of the histogram cannot reintroduce it without failing here.
+        """
+        labels, _ = pair_with_ignore()
+        counts = FN._label_counts_per_sample(labels)
+        for b in range(labels.shape[0]):
+            for p in counts.sum(dim=0).nonzero().flatten().tolist():
+                with self.subTest(sample=b, label=p):
+                    self.assertEqual(int(counts[b][p]), int((labels[b, 0] == p).sum()))
+
+    def test_ignore_voxels_are_not_counted_as_background(self):
+        labels, _ = pair_with_ignore()
+        counts = FN._label_counts_per_sample(labels)
+        # The structure is partly overwritten by the ignore block, so the foreground
+        # count is read back rather than assumed.
+        foreground = int((labels[0, 0] == 1).sum())
+        self.assertEqual(int(counts[0][0]), VOXELS - foreground - N_IGNORE)
+        self.assertEqual(int(counts.sum()), labels.numel() - N_IGNORE * labels.shape[0])
+
+    def test_subdivide_survives_the_ignore_label(self):
+        """The call that used to raise. `n_iters=1` keeps it quick -- the index is
+        drawn before the first iteration, so the path under test is covered."""
+        labels, image = pair_with_ignore()
+        fine, gen, out = FN.em_subdivide_labels(
+            image,
+            labels,
+            n_foreground_clusters=2,
+            background_clusters_range=(3, 3),
+            background_label=0,
+            n_iters=1,
+            max_fit_voxels=MAX_FIT,
+        )
+        self.assertEqual(fine.shape, labels.shape)
+        self.assertEqual(len(gen), len(out))
+        self.assertTrue(set(out).issubset({0, 1}), out)
+        # The ignored voxels keep generation label 0, which merges back to the
+        # background parent: they are treated as background, as they were before.
+        self.assertEqual(int(fine.reshape(labels.shape[0], -1)[0, :N_IGNORE].abs().sum()), 0)
+
+    def test_all_ignore_patch_returns_one_region(self):
+        """A patch with no annotated voxel at all resolves instead of raising."""
+        labels = torch.full((1, 1, *SHAPE), -1, dtype=torch.long)
+        image = torch.rand(1, 1, *SHAPE)
+        fine, gen, out = FN.em_subdivide_labels(image, labels, background_label=0, n_iters=1, max_fit_voxels=MAX_FIT)
+        self.assertEqual(gen, [0])
+        self.assertEqual(out, [0])
+        self.assertEqual(int(fine.abs().sum()), 0)