I went through several stages of AF doubts with a K10D. First I thought the AF was a bit wonky "sometimes", then I thought there were mirror alignment issues, because I had problems with MF too. Then I figured eyesight/dioptre issues, and, finally, I got around to noting whether the SR was on or off when shooting. Whenever I've tried to check for basic focus accuracy at around one to two metres, though, it's seemed spot on.
Since first noting problems, I've learned that the viewfinder system has its own effective aperture, making fast lenses look like they have more DoF than they do. Moreover, noting my technique and trying to find variables, I've discovered that with the K10D VF at least, my eye being off-axis also affects my impression of focus. All other things being equal, I find focus is pretty well always reliable and good with long lenses, even fast ones, the most trouble I've had has been with a 20mm f/1.8, results seem a bit variable with 50/1.4.
My suspicions have also fallen on the SR from time to time, not because the shots are motion-blurred rather than OOF (I know the difference) but because I thought I was seeing inconstencies even with an old MF lens which hadn't moved between shots. The wider & faster the lens, the worse it is.
To be able to move for shake-reduction, the sensor has to be free to move, obviously, and with SR on, it must be free to move during the exposure. Which implies that there's some play in it, however slight. Moreover, depth of focus (i.e. the depth in focus at the sensor plane, behind the lens) is tighter for wider apertures, obviously, but also tighter (at short distances) for wide angle lenses than for longer lenses (sounds paradoxical, but it's the opposite way around from depth of field in front of the lens - less relevant for distant subjects though.)
So wondered, could the focus accuracy of these moving-sensor cameras be determined somewhat by where, exactly the sensor is, and how sloppy it is in its sliding mountings, if it has sliding mountings? Is it magnetically clamped in place when SR is turned off? Does it wear and become sloppier with age? I have no idea, but I do *seem* to get more consistent (focus) results with SR off. I'm not sure whether it's changing between times if I switch SR on and back off again, though. At 20mm and f/1.8 and 12ft focal distance, I make the depth of focus at the film plane of the order of a thousandth of an inch, so the sensor hasn't got to be too loose to be problematic, but then I'd expect a lot less that 1/1000" play for all that. So, more tests needed :-)
If things are overall simply inconsistent then it's kind of hard to pin down, when testing stuff, whether whatever you're changing from shot to shot is having any effect.
I think for AF, there are other effects which might be creeping in below the radar, as it were - the colour and brightness thing has been mentioned, the focus detect does seem to be sensitive to light level (unsurprisingly) and also colour temperature, then there's the business of the VF indicators not necessarily being aligned with the actual effective focal points - I can see that this could quite easily be very different from lens to lens, and could in itself account for a lot of the inconsistencies seen in real-life situations. And I wonder how the focus sensors are sampled, and how often, because many light sources are changing brightness and colour quite dramatically at various frequencies. Old-style striplights, TVs and CRT screens, for example, flicker at a fairly low frequency and I wouldn't be entirely astonished to find that the AF system could be upset by this kind of thing.
As I understand it (which isn't very much) phase detect AF works in effect like a rangefinder with the lens diameter as its baseline, using linear CCDs to compare images from either side for coincidence (see this, for example:
Autofocus: phase detection ) - seems to me that if the light has changed somewhat between when the camera's CPU "looks" at the one AF sensor and the other, then it's not necessarily going to come up with the right answer. If there's this kind of effect going on, then "cross" type sensors (with pixels in two directions) are probably going to behave differently to "line" type ones. Moreover, if the focus system's optics suffer from any kind of chromatic aberration, then light (and subject) colour will, I imagine, be highly significant, and I further imagine that this will be affected by any lens chromatic aberration, which tends to be worse the further off-axis one gets, and worse with wideangles.
After pondering this for some time and imagining all the things which could affect image focus or confuse the focus detect system, I've kind of come around to thinking that its not too far short of miraculous that it works at all, and that I get more decent sharp pictures than I used to, and nothing's perfect, so perhaps I should stop worrying too much about it and just take photos :-)