VoxelForge/Standalone/Research/WorldAssemblerM3/Research15/REPORT.md

R15 — Whole-surface artifact audit

Key recommendations

1. Scan every LOD0 triangle in the exported region using three area-weighted metrics over overlapping surface patches: detrended geometric residual, smoothed normal variation, and periodic-band coherence. Group adjacent anomalous patches into clusters before failing the mesh.
2. Calibrate each scale against plain native terrain exported by the same mesher at 25 cm. Choose thresholds on held-out clean regions, then lock them. The gate should pass clean references and reject each of the five pre-fix defect meshes; if one slips through, improve the detector rather than tune the threshold around that mesh.
3. Add the offline check after mesh export in verify.py’s medium capture-and-audit stage. Report cluster coordinates and scores, plus a false-color mesh or marked capture. Full tier already reuses passed medium evidence; fast would need a small LOD0 export to include this gate.

Metrics

For triangle (f=(v_0,v_1,v_2)), compute area (A_f=\frac12|(v_1-v_0)\times(v_2-v_0)|) and geometric face normal (n_f) from its positions. Use these normals for shape checks; stored smooth MC normals can hide or exaggerate geometric changes.

• Detrended residual (D_r): Fit a robust local quadric (q_r) over a geodesic neighborhood of radius (r), and measure the area-weighted RMS signed normal distance: [ D_r(P)=\sqrt{\frac{\sum_{v\in P}A_v,\delta_r(v)^2}{\sum_{v\in P}A_v}}, \qquad \delta_r(v)=(v-q_r(v))\cdot n_r(v). ] Also retain the 95th percentile and largest adjacent residual jump in the patch. This measures relief and steps without assuming the surface is a floor or wall.

• Smoothed normal variation (G_r): For adjacent faces (f,g), let (\theta_{fg}=\cos^{-1}(n_f\cdot n_g)) and (d_{fg}) be their geodesic center distance. Average normals over radius (r) first, then compute an area-weighted RMS of (\theta_{fg}/d_{fg}). This catches clustered creases and teeth while reducing single-triangle MC facet noise.

• Periodic coherence (C_\lambda): Take a spectrum or autocorrelation of (\delta_r) across overlapping surface patches, in all local tangent directions. Score the strongest narrow-band energy relative to neighboring frequencies, and require phase coherence across adjacent patches. Search wavelengths from roughly two to 32 voxels (0.5–8 m); use multiple window sizes so the longest periods have enough support. This detects repeated ribs, washboard, and seams without relying on a chosen rail.

Use all three as a small, interpretable suite. A raw dihedral histogram loses location and mixes ordinary MC facets with real creases. Discrete Gaussian curvature (K_i=(2\pi-\sum_j\theta_{ij})/A_i^*) and cotangent-Laplacian mean curvature are useful visual diagnostics, but raw one-ring estimates can be noisy on irregular MC triangles; prefer multi-scale smoothing or quadric fits. Triangle quality, such as (2r/R) or minimum altitude-to-edge ratio, is a useful secondary mesh-health check, not a primary artifact detector: ridges and staircases need not contain slivers.

MC linearly interpolates grid-edge crossings; the output can miss undersampled details, contain poor-shaped facets, and fail to preserve sharp features. That makes universal angle, curvature, or triangle-quality cutoffs unreliable. CGAL’s isosurfacing manual and Dietrich et al. on MC triangle quality document these limitations. MeshLab and libigl provide useful curvature and quality measures; libigl also documents quadric curvature fitting. The specific three-metric combination above is a recommendation inferred from those methods and VoxelForge’s regressions, not a published universal detector.

Reference and threshold method

Build a reference set from native-only LOD0 meshes using the same 25 cm mesher, recipes, and export path. Match patches by surface orientation and broad shape class; compare patch scores, not individual triangle scores. Use a held-out clean set to set per-scale thresholds from the distribution of per-export maximum cluster scores (with block bootstrap or equivalent uncertainty), so scanning more surface does not silently increase false alarms.

Threshold a connected cluster of adjacent anomalous patches, with a minimum footprint calibrated to the smallest visible defect. Ignore isolated single-triangle excursions unless a separate topology or fragment check catches them. Lock thresholds before scoring candidate fixes. Require every pre-fix defect mesh to fail individually and clean native references to pass. Keep all five bad meshes as fixtures; the native holdout guards against thresholds that simply reject rough terrain.

Defect-to-metric map

Known defect	Primary detector	Evidence and expected signal
L35 wall ribs	(C_\lambda + G_r)	The old wall has measured periods near 4.75 m in normals and 7.44 m in radius. A coherent band should cluster over the wall area. L35 report
L36 floor washboard	(C_\lambda + D_r)	The old floor repeats at 1.438 m and 7.250 m, with 2.25–2.48 cm amplitude. Whole-surface patches include the tread that the wall-only scan missed. L36 report
L45 segmented ramp	(C_\lambda + D_r)	Segment seams create repeated surface residuals along the ramp. The L45 control already fails its floor-ripple gate at 17.9 mm against a 15 mm limit. L45 report L46 report
L46 entrance staircase	(D_r), including adjacent residual jump; (G_r)	The first 7 m measures 39.9–51.4 mm ripple and 81.8–82.9 mm residual steps. Detrending against the sloped floor separates the steps from the intended grade. L47 report
L47 wall-foot sawteeth	(G_r + D_r), clustered	Scan all connected faces, including the wall foot. The existing three floor rails do not reach it. L47 report

This three-metric set should catch all five while passing smooth native terrain if the matched reference and control gate confirm the separation. The reports establish the defects and measured signals; they do not establish final mesh-wide thresholds.

Implementation and reporting

The existing .vfmesh readers expose tile LOD, voxel size, positions, normals, indices, and masks. Extend the tile-streaming approach used by the ring audit and the tile scanner: analyze the whole target region at LOD0, include all face orientations, and weld duplicate vertices at same-LOD tile boundaries. Keep the audit offline on exported geometry.

The natural insertion point is after CaptureLandmarkL47.py and passage mesh exports in run_captures_and_mesh_audits(). That stage runs in medium and is reused by full; fast currently has no equivalent capture export. A fast-tier version would need a small targeted LOD0 export. Voxel Tools’ smooth-terrain documentation notes that voxel LOD stitches can still show sharp steps, while SAGA’s Terrain Ruggedness Index illustrates the value of neighborhood-scale terrain measures; its DEM assumptions do not cover arbitrary tunnel surfaces.

For each cluster, report world-space centroid, bounds, geodesic area and extent, triangle IDs, strongest metric, scale or wavelength, value, threshold, LOD, voxel size, and mesh/layout identity. Emit JSON plus a compact CSV, and a false-color mesh or marked image with cluster IDs and a reproducible camera pose. MeshLab’s face-quality coloring is one available review path.