REFERENCE SELF-VALIDATION — mvs_mesh_reconstruction_usd This is 60_validate_reference.py run against the shipped object_model.usdc. It shows the reference satisfies every criterion the submission publishes, and that the grader is runnable. It is NOT a blind trial and does NOT measure task difficulty — see 03_SHORT_FIELDS.txt, Difficulty Self-Test. Geometry (section C) is scored against the DENSE grader-only ground truth (820,000 points). metrics.json (section D) is scored against the 300-point sparse reference the solver receives. Three gates recompute from the mesh what earlier versions read out of metrics.json: A6e (UV overlap, exact polygon clipping), B7 (self- intersections, Moller-Trumbore), and B8 (the submitted self-intersection count must be a true int AND match B7). A6e is why the run below reports 77 overlapping UV pairs where metrics.json claims a flat 0.0 -- the submitted figure is a rasterised estimate. 0.000012% of UV area passes a 0.10% threshold, but the rubric now measures it instead of trusting it. reference: accuracy=0.2609mm completeness=0.1446mm normals=0.9967 tris=249820 PASS A1 parse+crate — crate=True type=Mesh PASS A2 single mesh, records>0 — mesh_prims=1 points=124912 tris=249820 PASS A3 required fields — normals=True uvs=True extent=True PASS A4 index integrity + all referenced — range_ok=True all_tris=True unreferenced=0 PASS A5 stage metadata — metersPerUnit=1.0 upAxis=Z PASS A6a UV primvar named exactly 'primvars:st' — name=primvars:st PASS A6b UV interpolation is faceVarying or vertex, length matches — interp=faceVarying n_uv=749460 loops=749460 verts=124912 PASS A6c every face carries UV, all finite — faces_without_uv=0 finite=True PASS A6e UV charts do not overlap (independently recomputed) — overlap_area=0.000012% of UV area, overlapping_pairs=77 (threshold 0.10%; metrics.json claims 0.0) PASS A6d normals cover the mesh and are unit length — interp=faceVarying n=749460 max_unit_dev=1.93e-07 PASS A7 finite geometry — no NaN/Inf in points, normals, st PASS B1 watertight (0 boundary edges) — boundary_edges=0 PASS B2 manifold — non_manifold_edges=0 vertex_manifold=True PASS B3 single component — components=1 PASS B4 orientable, volume>0 — orientable=True signed_volume_m3=0.002670 PASS B5 triangle band — triangles=249820 PASS B6 no degenerate tris — degenerate=0 min_area=3.815e-12 PASS B7 no self-intersections (independently recomputed) — self_intersecting_triangles=0 from 110383 candidate pairs (metrics.json claims 0) PASS B8 submitted self-intersection count is an integer and agrees — type=int value=0 recomputed=0 INFO B9 genus — euler=2 genus=0 PASS C1 scale <=3% — max_abs_extent_error=0.4266% PASS C2 completeness — mean=0.1446mm within1mm=99.84% PASS C3 accuracy — mean=0.2609mm within1mm=99.97% PASS C4 chamfer <=1.0mm — chamfer=0.2028mm PASS C5 normal consistency — mean_abs_cos=0.9967 within30deg=99.77% PASS C6 centroid <=2mm (closed form) — centroid_offset=0.6605mm PASS D1 valid json — 10 top-level keys PASS D2 required fields present — missing=[] PASS D3 internal consistency (every published number recomputed) — 21/21 agree PASS D4 units metres — units=metres PASS E1 projects into all views — 60 views, worst in-frame=1.0000, failing=0 PASS E2 not a copy of the sparse reference — exact-coordinate matches=0.000% PASS E3 resolves detail beyond the sparse reference — triangles=249820 vs 300 reference points (ratio 832.7x) PASS E4 format — USDC + metres 34 of 34 entries pass, 0 failed (33 gated, 1 informational). overall_pass=True Independent corroboration (not part of the gates above): - every ground-truth vertex projects inside the rendered alpha matte in all sampled views: 100.000% containment against a 2px dilation, <=2px bbox error - COLMAP registered 60/60 images at 0.233 px mean reprojection error with intrinsics AND extrinsics held fixed - visibility audit: 9.76% of the ground-truth surface is unseen by every camera, and 100% of that is the flat base - the cloud-meshing shortcut, run on the exact 300-point sparse reference extracted from input_materials.zip, scores 1.239mm accuracy / 51.6% within 1mm and fails ALL FIVE gates (C2 mean, C2 pct, C3 mean, C3 pct, B5 band). See shortcut_attack_verification.txt and grader/sparse_design_report.json.