Machine cuts smooth small circle but large circles are rough

I'm trying to figure out a problem with my Mill in that small circles are cut fine but large circles are 'faceted' slightly, like the machine is doing loads of straight lines. I'm using G2 and G3 codes so it should be doing arc's. It also made elipses instead of circles but that is nearly sorted, explanation follows.

A bit of research suggested stepper motor speed and acceleration. My velocity is set to 33mm/s and in the start acceleration was 1000mm/sec squared. I reduced the acceleration down to 100 mm/sec squared and that sorted the elipse, the discrepancy dropped from .7 mm to .2mm.

I've just changed from slides to guides and set the machine up within microns so I don't think it's a mechanical issue.

So I wondered if anyone else had experienced similar problems and if they could suggest how to fix them.
I've attached screen shots of the config file and a video showing the issue.

If you don't have something like G64 P0.005 in your gcode, this could help improve contour accuracy.

Odd how the flats appear on the top half of the disk edge (0:42 of your vid for example) but not lower down. Are you sure it's not something like the cutter vibrating/resonating in the cut as it works the material? Perhaps moving too quickly and/or cutting too deeply in one go?

Your video at 0:53 shows some of the G-code, but I can't see where you've used G2 or G3 - looks more like short straight lines to me.

The final cut I made wasn't at full depth. I had been playing around manually.
The final code was only two lines long, a g2 full semi circle and a further g2 full semi circle. I think it's to do with speed and acceleration on the steppers as adjusting that made this happen. Without adjustment it cut a fairly smooth elipse.

would this G64 code go before or after the G2?

G64 P0.005 or whatever P value you want, needs to be before G2 or G3. Then it is active, until reset or M2.
Did you check the fixed bearing clearance, or spindle backlash, because of the elliptical, it sounds more like a mechanical issue.
As a quick test to exclude this possibility, i would set a dial gauge and push pull on the involved axes.

I'd actually just done that before your reply and decided to go direct drive and beef up the mounting bracket. Backlash was .06 and flex on the bracket gave plus and minus 0.5 While the Y axis was moving you can see rhythmical flexing of the mounting bracket so I suspect this is the main culprit.
The ellipse was sorted a lot by changing the acceleration, basically the machine can't keep up with it's brain and misses step. Slowing accel reduced the out from .7 to .2 now once I have the bracket sorted I hope that drops down further and I'll re-evaluate.
I've just changed to linear guides so know that setup is within microns, I spent three days with the micrometers and gauges getting her perfect.