Just to clarify after that comment, my point wasn’t about approximate values but approximation of curves and surfaces which polygonal modeling is. There are no curves, just ones that are approximated with straight edges - More precision, more vertices and edges, more difficult to control and if you somehow manage, it’s still an approximation.
In 2D/3D CAD (NURBS or Solid Modeling) one could define a circle and the parameters could be a position and a radius (and height for a 3D cylinder). It gives a nice circle that has the same radius all around no matter which point is measured. The precision is given.
In polygonal modeling we add a circle, give it a position and a radius, and the number of vertices/sides. The result is an approximation of a circle which has that radius but only at the points where the vertices are. We could start with more vertices and get better approximation, but then we would also increase the complexity right from the start and have to control more geometry while modifying this shape.
Subdivision surface is a popular method to use a lot of geometry in a model while still maintaining controllable amount of control points. Most used subdivision algorithm across different packages is Catmull-Clark which Blender also uses, and it’s an approximation algorithm instead of interpolation. This gives nice looking curves but as precision goes, it’s even worse because it’s based on the control cage which is an approximation, and it then approximates that.
So the precision depends on the control cage and if we want less control points and use high subdivision level, it gives much smaller version of the original circle that had radius (or diameter) specified at the start. The subdivision doesn’t happen uniformly either, the diameter is not the same all around.
That’s still just a circle, not a complex 2D or 3D shape. And none of that usually matters since we don’t need absolute precision when creating polygonal models with polygonal modeling tools. The precision is relative, relative to: features on the object itself, other objects, distance to camera, scale, detail level, and even render resolution.
Common instruction is to “have the finalized model in real world size” but forget to mention that within 1-3% is good enough when selling models. 3D printing almost counts because curves and surfaces are defined only with real geometry and we need to think about printing tolerances, clearances for interlocking parts, etc. which is a pain in a backside. Luckily those tolerances and clearances are still quite loose, smallest being 0.1mm which in engineering is like waving a finger in a hangar.