No, I don't think that it is a tool compensation problem as the gcode preamble includes G40 which should turn it off. That amount of error is strange - too big for cutter diameter error, I would have thought, but too small for any significant axis calibration error (which I have checked in the past). The error is also pretty much the same, within the limits to which you can measure MDF anyway, along both axes, which I would have thought meant that it's probably not machine bend as I would expect that to vary with axis. However, this is my MDF router, so anything's possible. At least the error is consistent enough that I can tune it out by tweaking cutter diameter in vCarve. Sometimes I have to pretend to be an engineer and not a scientist (it works, dammit, so forget the theory!).
Thanks for the thoughts, guys, and I shall try putting a mike across the plain part of the cutter (3mm single flute on 6mm shank) as a check.
This is what I'm playing with.
Click image for larger version.  Name: Cable chain small.jpg  Views: 407  Size: 80.9 KB  ID: 11750
It's a cable chain idea based on something I found a while back; it needs reasonably accurate cutting to make it all stay together (no glue or loose fixings) while still being able to articulate.