In your 10,000 "steps" are you including micro stepping by the driver - maybe 8 or 10 or even 16 microsteps. Stepper motors do not stop accurately on microsteps.
However to echo Chip - this is backlash.
Your problem is in your fine tuning. You have tuned this problem into the system.
Back lash is the non-movement of the axis - and from what you say appears to be about 7 thou on your system. That 7 thou is a constant on any change of direction - and should be taken up before the axis moves.
There should be no need to "fine tune" an axis drive - it is a mathematical formula - x number of steps per rev of motor - y number of revs per 1 inch of movement - answer x times y. (Mine is 60,000 - if you are using a metric screw for 1 inch of movement the result is not as neat - but is still calculable)
What you have done is fine tune out the backlash and shared it along 1 inch of movement, so you are actually moving maybe 0.1007 for every 0.1 of an inch. Move 10 times for the inch and this gets rid of your backlash. However - for the first tenth, there is still a full 7 thou of backlash -maybe nearer 8- and you have only compenstated for 0.7 of a thou. On the second tenth you have compensated a bit more, but you are still short - and so on one up to 1 inch where you are accurate.
What you must do is calculate your steps per inch - and put this in your motor tuning.
Check then to see if there is any backlash. Move the axis to the right, Stop, Setup your dial test indicators and Zero them. Move to the right a further inch using a G0 X1 command on the MDI line and then move back using a G0X0 command. Check your DTI and the result is the backlash - that distance the axis doesn't move as the gears etc, settle in to push the other way. (You dont have to actually move as far as an inch - backlash is not dependent on distance - half and inch or quarter would do - as long as the movement is equal.)
7 thou actually seems pretty good - I would hate to tell you what mine is - and you may be able to adjust it out mechanically - but despite what everybody says, there will still be some. If everything was so accurate - nothing would move - it might, however, be so small as not to be noticeable or measurable.
Enter this amount in the backlash table and turn the backlash on. You should then get accuracy all along the range, because the backlash is taken up first by Mach3.