Picking up Jonathan's comment re stiffness in roll and manufacturer's tolerance...

HiWin 20mm rails have a maximum permitted height difference of 0.130mm with 500mm rail spacing. That corresponds to a difference of 0.013mm across that 50mm or so bearing block. Don't know about you, but that's about a quarter of my thinnest feeler - and that's the maximum allowable tolerance. As for stiffness - and I don't recommend you do this at home as I only did it by accident - my gantry structure probably weighs around 30kg, and I managed to leave it hanging while only attached to the carriage at one end. It probably dropped by no more than 25mm or so. I hate to think of the abuse of the pair of bearing blocks taking the weight, but what's done is done. I shall be reporting on my experiments with epoxy (i.e. mistakes and how I've tried to correct them) elsewhere when I'm actually happy with the results, but in simple terms, it's a lot easier to get the rails level than to remove twist. The main reason to get the rails level (given that you will take out minor variations by machining the spoil board anyway) is that with, say, a 1000mm rail spacing, the max rail height difference spec only allow a height difference of about 0.26mm (10 thou). That's not too difficult with epoxy, and I would expect rather better with just a bit of care. But twist? As long as it is constant, then the gantry to bearing block joint can be shimmed/adjusted to suit but you only need a tiny amount of varying twist to go way out of spec. I'm still trying to get mine somewhere near acceptable. I'm using a sensitive level (0.02mm/m readings) to check.

It would be easy to say that the problem is that I'm chasing irrelevant numbers with tools that are working at their limit, except that a trial bolt-down of rails with gantry mounted gave clear signs of graunching bearings even when I was happy with levels. Suck-it-and-see tells me that there's something wrong.