diff --git a/delaynet/network_analysis/_normalisation.py b/delaynet/network_analysis/_normalisation.py
index 5a0dc44..16e7aaa 100644
--- a/delaynet/network_analysis/_normalisation.py
+++ b/delaynet/network_analysis/_normalisation.py
@@ -152,11 +152,25 @@ def wrapper(
         x_true = metric_fn(weight_matrix, *args, **kwargs)
 
         # Sample ensemble and compute metric values
+        # Retry if the metric raises ValueError on a random graph (e.g., reciprocity
+        # rejects accidentally symmetric matrices). Cap attempts to avoid infinite loops.
         samples = []
-        for _ in range(n_rand_val):
+        max_attempts = n_rand_val * 10
+        for _ in range(max_attempts):
+            if len(samples) >= n_rand_val:
+                break
             R = _random_directed_gnm_igraph(n, m)
-            x_r = metric_fn(R, *args, **kwargs)
+            try:
+                x_r = metric_fn(R, *args, **kwargs)
+            except ValueError:
+                continue
             samples.append(np.asarray(x_r))
+        else:
+            raise RuntimeError(
+                f"Could not collect {n_rand_val} valid null-distribution samples "
+                f"after {max_attempts} attempts. The metric {metric_fn.__name__!r} "
+                f"raised ValueError on every random graph generated."
+            )
 
         samples_arr = np.stack(samples, axis=0)  # shape: (n_random, ...)
         mu = samples_arr.mean(axis=0)
diff --git a/tests/network_analysis/test_normalisation.py b/tests/network_analysis/test_normalisation.py
index f3806dc..30f54a7 100644
--- a/tests/network_analysis/test_normalisation.py
+++ b/tests/network_analysis/test_normalisation.py
@@ -8,6 +8,8 @@
 - Seed reproducibility for z-scores.
 - Sigma==0 branch returns 0.0 without NaN/inf.
 - Output shape/type parity when normalising.
+- Retry when a random graph raises ValueError (e.g. symmetric for reciprocity).
+- RuntimeError when no valid null-distribution samples can be collected.
 """
 
 from __future__ import annotations
@@ -18,6 +20,8 @@
 import pytest
 from igraph import Graph
 
+import delaynet.network_analysis._normalisation as _normalisation
+
 from delaynet.network_analysis.metrics import (
     link_density,
     reciprocity,
@@ -302,3 +306,44 @@ def test_example_matrices(network_metric_and_kind, n, m):
             assert not np.isnan(val)
             assert not np.isnan(z)
             assert np.isfinite(z)
+
+
+@pytest.mark.parametrize("n_rejected", [1, 3])
+def test_normalisation_retries_when_random_graph_rejected(monkeypatch, n_rejected):
+    """Samples that raise ValueError (e.g. accidentally symmetric for reciprocity)
+    are resampled until enough valid samples are collected."""
+    # Arrange
+    A = np.array([[0, 1, 0], [0, 0, 1], [1, 0, 0]], dtype=int)
+    symmetric = np.array([[0, 1, 1], [1, 0, 0], [1, 0, 0]], dtype=int)
+    partial = np.array([[0, 1, 1], [1, 0, 0], [0, 0, 0]], dtype=int)
+    independent = np.array([[0, 1, 0], [0, 0, 1], [1, 0, 0]], dtype=int)
+    generated = iter([symmetric] * n_rejected + [partial, independent])
+    monkeypatch.setattr(
+        _normalisation, "_random_directed_gnm_igraph", lambda n, m: next(generated)
+    )
+
+    # Act
+    z = reciprocity(A, normalise=True, n_random=2)
+
+    # Assert
+    assert isinstance(z, float)
+    assert np.isfinite(z)
+
+
+@pytest.mark.parametrize("n_random", [1, 2])
+def test_normalisation_raises_when_all_random_graphs_rejected(monkeypatch, n_random):
+    """If every sampled graph raises ValueError, a RuntimeError is raised instead of
+    building an incomplete null distribution."""
+    # Arrange
+    A = np.array([[0, 1, 0], [0, 0, 1], [1, 0, 0]], dtype=int)
+    symmetric = np.array([[0, 1, 1], [1, 0, 0], [1, 0, 0]], dtype=int)
+    monkeypatch.setattr(
+        _normalisation, "_random_directed_gnm_igraph", lambda n, m: symmetric
+    )
+
+    # Act / Assert
+    with pytest.raises(
+        RuntimeError,
+        match=f"Could not collect {n_random} valid null-distribution samples",
+    ):
+        reciprocity(A, normalise=True, n_random=n_random)