Well, I guess better to have it resolution dependent, and best to have it local per bond or so. I have no intention to open another can of worms, but where those 0.02 and 2 come from? Can you really see that level of details at say 2.5A resolution and lower? I guess at that resolutions they should be both zero. In contrast, at higher resolution larger deviations are justified just by the data and so larger margins in restraints should be allowed. I know there is a heap of good papers on this matter, including the recent ones, and there are long threads on bb, but to me this question is still somewhat open. May be someone on the board can thoroughly/convincingly comment on this, summarizing all previous discussions?
My guess would be that the limits are derived from Engh & Huber. I was taught to use 0.016 and 1.6; 0.02 and 2 are a bit on the high end but for very high-resolution structures this wouldn't be inappropriate, and it's arguably helpful earlier in refinement to let the geometry be a bit looser. I don't think it's practical to force them to be zero at high resolution - it will make it much harder to converge during refinement. Regardless of what the limits *should* be at different resolutions, reviewers will expect something below 0.02/2, and the overwhelming majority of protein structures will be done at resolutions where these values are indeed appropriate at upper limits.
-Nat