diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6ea8cc3..39e74b5 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -526,6 +526,15 @@ and `rbd._sensitivity`'s `levers` and `Lever` are gone (use `levers()`,
 
 ### Fixed
 
+- **A fixed maintenance duration is fixed whatever numpy's rounding.** A
+  duration known exactly (an `ExactEventTime`) is followed as a fixed
+  time, where a spread one is convolved; it was told from its variance,
+  a difference of two numbers about the duration squared, against a
+  tolerance a few units of rounding wide: on numpy 2.0 the variance came
+  out at 1e-15, and the exact and numerical analyses of a component
+  maintained for a fixed time refused ("its later age replacements ...
+  are too many ... to follow exactly"). The tolerance is a millionth of
+  the duration.
 - **A shut at an offset off the grid keeps its period (#319).** The
   grid's step for `NonRepairable.optimal_calendar_policy` was sized to
   put every period, offset and date on it, and fell back to the
diff --git a/repyability/rbd/_point_availability.py b/repyability/rbd/_point_availability.py
index 37223be..dd94b58 100644
--- a/repyability/rbd/_point_availability.py
+++ b/repyability/rbd/_point_availability.py
@@ -831,7 +831,10 @@ def __init__(
         mean, variance = _moments(
             sf, np.linspace(0.0, self.top, 4097), self.splits
         )
-        if variance <= (1e-9 * self.top) ** 2:
+        # A fixed time, to rounding: the variance is a difference of two
+        # numbers about ``top ** 2``, so it is uncertain by a few units of
+        # rounding in them (numpy 2.0 sums them to 1e-15 at a top of 5).
+        if variance <= (1e-6 * self.top) ** 2:
             self.fixed = mean
             return
         from scipy.signal import fftconvolve
diff --git a/repyability/tests/test_point_availability.py b/repyability/tests/test_point_availability.py
index 195e7cb..87353fb 100644
--- a/repyability/tests/test_point_availability.py
+++ b/repyability/tests/test_point_availability.py
@@ -491,7 +491,9 @@ def test_matches_the_simulation(name):
 def test_instant_repair_is_always_up():
     rbd = alone(unit(W([100.0, 2.0]), "instant"))
     assert rbd.point_availability([0.0, 10.0, 1e4]).tolist() == [1, 1, 1]
-    assert rbd.mission_availability(100.0) == 1.0
+    # To rounding: the integral of a constant 1 sums the quadrature's
+    # weights, exact to a few units of rounding in some numpy versions.
+    assert rbd.mission_availability(100.0) == pytest.approx(1.0, abs=1e-15)
 
 
 def test_units_dead_on_arrival_and_units_that_never_fail():