pub const STATED_PRECISION_INCONSISTENCY_RATIO: f64 = 10.0;Expand description
How far the measured deviation may exceed the stated precision before the file is called inconsistent with itself.
A decade, set from the corpus (54 files). Of the 48 that state a
precision, 25 state exporter boilerplate (1.E-17 m = 1e-14 mm, under
double resolution at any part size) and measure 3e4× to 5e14× over it; 14
state a real export tolerance their geometry meets, at 0.07× to 2.3× (two
outliers at 4.8× and 9.0×); and 9 state a real figure their geometry
misses by 12× to 5e4× (EngineBlock 1e-7 mm against 5.1e-3 measured;
raspberry_pi_3_molex 1e-5 against 9.8e-4). A decade sits in the gap
between the second and third groups, says “an order of magnitude” in the
same breath it names the numbers, and leaves headroom for what the
measurement itself cannot see below: the edge reading is the vendor’s 3D
curve against the kernel’s OWN projected pcurve, whose fit residual is of
order 1e-6..1e-5 mm, so a stated 5e-6 mm cannot be judged finer than that.