Originally posted by kh1234567890 It is not quite as simple as that. If you read the patents you'll realise how much cleverness goes into moving the sensor plate at the right velocity and direction during the exposure
Yes, I'm sure there's an awful lot of cleverness - the SR algorithms have to take rotational acceleration values from the shake-sensing accelerometers, and then, as you're saying, figure out directions and speeds. (That no doubt accounts for the delay in the green SR indication in the viewfinder.)
Assuming the camera's made the right deductions concerning rotational velocity, then the only other factor affecting how fast it moves the sensor is focal length. The amount of blur will be (1 - {value you input}/{actual FL})*{blur with SR off}. (A negative value just means the blur is in the opposite direction to the original blur.) So, obviously, you can end up with more blur than with SR off when the value you input is higher than the actual value. Conversely, for values lower than the actual FL, the blur you end can't ever be greater than the original blur.
If there is inaccuracy in the actual focal length of the lens, this will add to any error in the camera's deductions, and that's why it's always best to err on the low side when setting your FL value.