diff --git a/src/RioJoy.Core/Calibration/AxisCalibrator.cs b/src/RioJoy.Core/Calibration/AxisCalibrator.cs
index d05a708..be5ca99 100644
--- a/src/RioJoy.Core/Calibration/AxisCalibrator.cs
+++ b/src/RioJoy.Core/Calibration/AxisCalibrator.cs
@@ -200,7 +200,19 @@ public sealed class AxisCalibrator
             // 16838 (not 16383) and a +2 nudge are deliberate legacy anti-snap tweaks.
             _joystickXLast = lJx > 0 ? 16838 - ((lJx + 2) * sRightRate) : AxisOutputs.Center;
         }
-        // lJx == 0 leaves _joystickXLast unchanged (legacy behavior).
+        else
+        {
+            // Deliberate divergence from the legacy port, same reasoning as the
+            // throttle detent above: the legacy held the previous output on a raw
+            // of exactly 0. A real pot jitters and never rests at exact 0, so the
+            // hold was invisible on hardware - but vRIO's pad deadzone emits
+            // sustained exact zeros on release, and the hold latched the last
+            // in-motion output indefinitely: the ship kept turning at whatever
+            // rate the stick commanded the instant before release (bench
+            // 2026-08-01, Descent 3 drifting on yaw/pitch with the stick
+            // centered). Exact center in must be exact center out.
+            _joystickXLast = AxisOutputs.Center;
+        }
 
         return _config.InvertX ? AxisOutputs.Max - _joystickXLast : _joystickXLast;
     }
@@ -223,7 +235,12 @@ public sealed class AxisCalibrator
             lJy -= DeadzoneJoystick;
             _joystickYLast = lJy > 0 ? AxisOutputs.Center - (lJy * sDownRate) : AxisOutputs.Center;
         }
-        // lJy == 0 leaves _joystickYLast unchanged (legacy behavior).
+        else
+        {
+            // Exact 0 centers rather than holding the previous output - see the
+            // matching branch in JoystickX for the full story.
+            _joystickYLast = AxisOutputs.Center;
+        }
 
         return _config.InvertY ? AxisOutputs.Max - _joystickYLast : _joystickYLast;
     }
diff --git a/tests/RioJoy.Core.Tests/Calibration/AxisCalibratorTests.cs b/tests/RioJoy.Core.Tests/Calibration/AxisCalibratorTests.cs
index 7496c54..6189fbe 100644
--- a/tests/RioJoy.Core.Tests/Calibration/AxisCalibratorTests.cs
+++ b/tests/RioJoy.Core.Tests/Calibration/AxisCalibratorTests.cs
@@ -163,4 +163,31 @@ public class AxisCalibratorTests
         cal.Update(Report(throttle: -400)); // power-on with lever pushed: becomes the start
         Assert.Equal(0, cal.Update(Report(throttle: -100)).Z); // above max → re-based, rest
     }
+
+    [Fact]
+    public void Joystick_ReleaseToExactZero_Centers_NotHeld()
+    {
+        // Regression: the legacy left the output unchanged on a raw of exactly 0,
+        // which a jittering pot never produces - but vRIO's pad deadzone does,
+        // sustained, on every release. The held output latched the last in-motion
+        // value and the ship kept turning with the stick centered (bench
+        // 2026-08-01). A fast release goes straight from deflected to exact 0
+        // with no intermediate sample; both axes must land on center.
+        var cal = new AxisCalibrator();
+        cal.Update(Report(x: -3000, y: 2000)); // hard over
+        AxisOutputs o = cal.Update(Report(x: 0, y: 0)); // released, deadzoned to exact 0
+        Assert.Equal(AxisOutputs.Center, o.X);
+        Assert.Equal(AxisOutputs.Center, o.Y);
+    }
+
+    [Fact]
+    public void Joystick_ExactZero_StaysCentered_AcrossRepeatedPolls()
+    {
+        // The latch fed forward: once held off-center, every subsequent exact-0
+        // poll (one per 55 ms, forever) kept it there. Center must be stable.
+        var cal = new AxisCalibrator();
+        cal.Update(Report(x: 5000));
+        for (int i = 0; i < 5; i++)
+            Assert.Equal(AxisOutputs.Center, cal.Update(Report(x: 0)).X);
+    }
 }