3D Contouring - Loss of Accuracy

Hi all,

I'm stumped! On my BP-clone knee mill, I've tried everything I can
think of and nothing seems to produce the results I'm looking for. What
I find is that running a large 3D file I gradually accumulate a small
error typically in one of the most active axes which often exceeds 0.3mm
or .012" after half an hour to an hour of running time. It is usually
in the same direction, but not always. Below are some of the things
I've tried:

Large Steppers couple directly via heli-cal couplings running G210
Geckos and Flaschcut CNC - 650VA PSU
Large Steppers couple directly via heli-cal couplings running G210
Geckos and Indexer LPT & G-Code Controller - 650VA PSU
Large Servos running G320 Geckos via 5mm HTD belts & reduction pulleys -
both Flashcut and G-Code Controller - 1500VA PSU.
Large Servos running Rutex R990's via 5mm HTD belts & reduction pulleys
- both Flashcut and G-Code Controller - 1500VA PSU

Even if I remove the couplings or belts (so the motors aren't loaded)
the error remains. The only thing that seems to work is slowing things
down to snail pace rapids, with very low acceleration rates. I don't
think it related to a lack of power in the motor - the stepper was an
1870 oz-in and the servo motors are 880 oz-in at 17amps (which my PSU
can supply) through 2:1 or 1.75:1 belt reductions. All axes have zero
backlash ball screws fitted with dual thrust bearings at one end.

I'm an EE and while I don't claim to know everything about noise and
driving optos etc. I do have a 100Mhz scope and have observed clean,
full level step & direction pulses going into the Gecko's/Rutex's etc.
The power supply wiring is neat and organised, with 36,000uF of caps
fitted to the output. All cables are shielded and grounded at one end.
All terminations are made either with soldered DB-25 connectors or
AMP/TYCO CPC connectors with crimped pins using the correct tool etc.
All servo encoders are running in differential mode.

I've tried both old and new versions of Flashcut and the current version
of G-Code Controller. These tests have been run over the last (almost a
year) and in each case, a similar magnitude error occurs. I know one
would then start to suspect my CAM program - Visual Mill but I run in
absolute mode and I can see the positions at the end of the test file
specifiying either machine zero or program zero and then using my limit
switches or long nails hotmelt glued to the pulley as a zero reference
when the controller zero's the table, the error is there...

If I run relatively simple 2D or 2.5D files the error doesn't seem to
occur - it is only a function of large multi-axis contouring files.
Some of these problem files have 10,000+ lines of g-code code -
typically creating a "point cloud" that the software then contours.
BTW, I'm not criticizing the abovementioned products - they work great
in everything except these large 3D files. I seeking suggestions for
what nugget of wisdom I'm still missing.

Is it possible that even a servo system can occasionally lose a step or
two per million? If so, could that be why commercial CNC machines use
glass scales mounted on each axis to provide an absolute position
feedback loop in addition to a (motor) shaft mounted encoder? I'm not
quite sure how to incorporate such a thing, but I think it may be
possible with G-Code Controller and Indexer.

Anyone with experience in CNC contouring of large 3D g-code files or how
commercial CNC machines are made - can you help with any suggestions?
Thanks.

Cheers, Peter