Abstract
Geometric validity and perceived quality are different properties of a generated mesh. We measured how far deterministic, CPU-only mesh probes (boundaries, components, self-intersection, hidden surface, thin walls, roughness) track each, using expert defect labels, human geometry ratings, MLLM pseudo-labels and defects injected into real scans, in pre-registered analyses. Injected holes and floaters were detected perfectly (AUROC 1.000); interpenetrating parts and crossing sheets at 0.940-0.986, hidden shells at 0.811-0.912, thin fins at 0.952 (two larger severities) and vertex noise at 0.753-0.999, although the hidden-surface and self-intersection checks also respond to other defects. On 3D-DefectBench expert-agreement cells, a logistic regression on 30 probe features trained only on crowd labels reached a Matthews correlation of 0.272 averaged over three primary geometry defects, against 0.335-0.341 for the best vision-language judges; no paired McNemar test against the four strongest judges was significant, although one bootstrap interval favoured Gemini 3.1 Pro, and form/surface performance was largely explained by generator identity. Assets that crowd workers judged defect-free were geometrically dirty, while strict checks fired on undamaged scans. Within generators, correlations with MATE-3D human geometry ratings had absolute value at most 0.2; on 2,797 Hi3DBench meshes, correlations with an MLLM Geometry Plausibility score were at most 0.111 in absolute value, did not survive adjustment for prompt, and reversed for a geometric-detail score. A training-free conjunction of vision-language votes and probe flags raised Matthews correlation in 29 of 36 pairs but was significant in only one. Such probes are useful validity gates and confound checks, not perceptual judges.