diff --git a/src/state_point.cpp b/src/state_point.cpp
index 608918cf428..ff9111a3cb0 100644
--- a/src/state_point.cpp
+++ b/src/state_point.cpp
@@ -368,7 +368,11 @@ extern "C" int openmc_statepoint_write(const char* filename, bool* write_source)
 void restart_set_keff()
 {
   if (simulation::restart_batch > settings::n_inactive) {
-    for (int i = settings::n_inactive; i < simulation::restart_batch; ++i) {
+    // k_generation has one entry per generation rather than per batch, so the
+    // active portion of the array must be indexed by generation
+    int i_start = settings::gen_per_batch * settings::n_inactive;
+    int i_end = settings::gen_per_batch * simulation::restart_batch;
+    for (int i = i_start; i < i_end; ++i) {
       simulation::k_sum[0] += simulation::k_generation[i];
       simulation::k_sum[1] += std::pow(simulation::k_generation[i], 2);
     }
diff --git a/tests/unit_tests/test_restart_keff.py b/tests/unit_tests/test_restart_keff.py
new file mode 100644
index 00000000000..7c5ad8f7d93
--- /dev/null
+++ b/tests/unit_tests/test_restart_keff.py
@@ -0,0 +1,67 @@
+"""Tests for resuming an eigenvalue simulation from a statepoint."""
+
+import re
+
+import numpy as np
+import pytest
+
+import openmc
+
+
+# One line is printed per generation, e.g.
+#       9/1    1.18342    1.18055 +/- 0.00287
+# capturing the batch, the generation and the running average k-effective.
+_GENERATION_LINE = re.compile(
+    r"^\s*(\d+)/(\d+)\s+[\d.]+\s+([\d.]+)\s+\+/-\s+[\d.]+\s*$", re.MULTILINE
+)
+
+
+@pytest.mark.parametrize("generations_per_batch", [1, 5])
+def test_restart_keff(run_in_tmpdir, capsys, generations_per_batch):
+    """A restarted run must resume with the correct accumulated k-effective.
+
+    ``simulation::k_generation`` holds one entry per generation rather than one
+    per batch, so the accumulation in ``restart_set_keff()`` has to be indexed
+    by generation. Indexing it by batch sums the wrong entries -- and too few of
+    them -- leaving the restarted run with a source-normalization k-effective
+    roughly a factor of ``generations_per_batch`` too low, which over-produces
+    fission sites and fills the shared fission bank.
+    """
+    model = openmc.examples.pwr_pin_cell()
+    model.settings.particles = 500
+    model.settings.inactive = 3
+    model.settings.batches = 8
+    model.settings.generations_per_batch = generations_per_batch
+
+    statepoint = model.run()
+
+    with openmc.StatePoint(statepoint) as sp:
+        assert sp.generations_per_batch == generations_per_batch
+        n_inactive = sp.n_inactive
+        n_loaded = len(sp.k_generation)
+
+    # Restart from the statepoint, asking for two more batches
+    model.settings.batches = 10
+    model.export_to_model_xml()
+    capsys.readouterr()
+    openmc.run(restart_file=statepoint)
+    output = capsys.readouterr().out
+
+    # Symptom of a low restart k-effective, asserted separately so that the
+    # failure mode is obvious
+    assert "fission bank is full" not in output
+
+    with openmc.StatePoint("statepoint.10.h5") as sp:
+        k_generation = np.asarray(sp.k_generation)
+
+    average_k = [float(m[2]) for m in _GENERATION_LINE.findall(output)]
+    assert len(average_k) == 2 * generations_per_batch
+
+    # The running average reported for each generation of the restarted run must
+    # be the mean over every active generation simulated so far, including the
+    # ones loaded from the statepoint. The comparison is exact to the precision
+    # of the printed value.
+    first_active = generations_per_batch * n_inactive
+    for i, reported in enumerate(average_k):
+        expected = k_generation[first_active:n_loaded + i + 1].mean()
+        assert reported == pytest.approx(expected, abs=1e-5)