ok - just got a response from hezza on cgsociety. here's his solution - 'Its a known problem with the Catmul Clark subdivision method that was introduced in 8.5, I really don't understand how anyone actually uses mental ray displacement in production without switching it back to the old division method If you type addAttr -ln "miExportCCMesh" -at bool mentalraySubdivApprox1; the problem will go away, but you loose all the memory/speed improvements that the CCmesh method is meant to give us. Please please report this as a bug to autodesk, its getting a joke this hasn't been fixed in 3 releases.' haven't tried it yet as i'm at work.