According to Randy Howard  <randyhoward@FOOverizonBAR.net>:

    [ ... ]

    I don't know of an article, but I'll tell you how I *think* that
they work.  Obviously, they need to derive power from the camera's body.

    Start with a set of sensors which detect angular changes in each
of the primary two axes on the film.  (There is nothing which the VR can
do about rotating the lens around its center axis, as far as I know.)

    Then add an optical element close to the nodal point of the
lens, which can be either slid up-and-down or sideways, or which can
tilt the optical element -- with either causing the image to move on the
sensor (or film).

    Finally -- add a bit of computation gear which can translate the
sensed motion in the two axes into control voltages applied to the
optical element so when the motion would make the image move on the
sensor, the controller adjusts the position of the element so it keeps
the image steady -- at least until you run out of travel in the element,
such as if you are pivoting to track a sports event or something
similar.  (I've not handled a Nikon lens with VR, but the one Canon lens
so equipped was supplied with a switch to turn off the VR when you are
tracking so it does not fight you.)

    Of course, zoom optics complicate the computation, as how much
you need to adjust the position of the image on the sensor would be a
function of the current focal-length setting.

    And the one camera which does it in the body instead of in the
lens does it (I believe) by moving the sensor up and down or sideways.

    Needless to say, it makes no sense for the VR to work full time,
so it needs to receive a signal from the camera that the shutter release
is half-depressed so it should start compensating.

    I would be interested to see replies by others who may know more
correcting my guesses.

    Can you tell whether the VR feature was turned on during the
trouble shots?  Could it be that you were panning without switching off
the horizontal correction, and your exposure happened during a time when
the lens was quickly snapping back to the other side of the travel to
try to stabilize again after running out of travel?  Does the camera
write whether VR was in effect in the EXIF data?  If so, you should be
able to tell when it was enabled and when it was not.

    I can understand that.  Does that lens even handle full frame?
From the notice on the B&H web page for that lens, there is the
following warning:

======================================================================
Important Notice!
Use of DX lenses are NOT recommended with 35mm or APS SLR cameras
======================================================================

which suggests that it does not properly fill the 35mm frame.

    That does indeed set limits.

    You might try turning on and off the VR feature, and see whether
it makes a difference with the shots you get -- trying to duplicate the
illumination angles which were in effect when you got the problem shots.

    Good Luck,
        DoN.