Summary
Union of two equal cylinders meeting at right angles returns a volume equal to 2 × cylinder instead of 2 × cylinder − Steinmetz_overlap. The kernel appears to leave both cylinder surfaces in the output without computing or removing their intersection region.
Surfaced by mecheval/tasks/a3-cross-shaft-01.json, where 23 / 31 runs across all 6 models in the leaderboard fail mass_props with the exact same wrong volume — strong signal that the model output is fine and this is a deterministic kernel bug.
Reproduction
{
"version": "0.1",
"nodes": {
"1": {"id": 1, "name": "vertical", "op": {"type": "Cylinder", "radius": 10, "height": 40, "segments": 32}},
"2": {"id": 2, "name": "horizontal", "op": {"type": "Cylinder", "radius": 10, "height": 40, "segments": 32}},
"3": {"id": 3, "name": "h_rotated", "op": {"type": "Rotate", "child": 2, "angles": {"x": 0, "y": 90, "z": 0}}},
"4": {"id": 4, "name": "h_pos", "op": {"type": "Translate","child": 3, "offset": {"x": -20, "y": 0, "z": 20}}},
"5": {"id": 5, "name": "cross", "op": {"type": "Union", "left": 1, "right": 4}}
},
"materials": {}, "part_materials": {},
"roots": [{"root": 5, "material": "default"}]
}Expected volume: 2·π·10²·40 − (16/3)·10³ = 25132.74 − 5333.33 = 19799.41 mm³
Actual volume: 24971.6 mm³ (kernel reports approximately 2× the cylinder volume — Steinmetz region is being double-counted)
Affected mecheval tasks
a3-cross-shaft-01 — 23 deterministic failures all reporting actual volume ≈ 24971.6, 0/31 pass-rate across Opus / Sonnet / Haiku / GPT-5 / GPT-5-mini / GPT-4o-mini
Probable root cause
In the SSI stage of the boolean pipeline, two perpendicular cylinders intersect along a pair of curves (two ellipse-like seams forming the boundary of the Steinmetz solid). If the SSI is missed or the cylindrical-face splits aren't classified, the trim/sew stage emits a result that contains both full cylinders' lateral surfaces overlapping in space — manifold is closed (each cylinder is closed on its own), but the volume integral over the unified mesh effectively double-counts the overlap region.
Worth a cargo test -p vcad-kernel-booleans --features debug-boolean test_perp_cylinder_union -- --nocapture run focused on the Cylinder × Cylinder SSI case in crates/vcad-kernel-booleans/src/ssi.rs.
Acceptance
Volume of cylinder(r=10, h=40, axis=Z) ∪ cylinder(r=10, h=40, axis=X, centered at z=20) returns 19799.41 ± 1%.
Refs
Summary
Unionof two equal cylinders meeting at right angles returns a volume equal to 2 × cylinder instead of2 × cylinder − Steinmetz_overlap. The kernel appears to leave both cylinder surfaces in the output without computing or removing their intersection region.Surfaced by
mecheval/tasks/a3-cross-shaft-01.json, where 23 / 31 runs across all 6 models in the leaderboard fail mass_props with the exact same wrong volume — strong signal that the model output is fine and this is a deterministic kernel bug.Reproduction
{ "version": "0.1", "nodes": { "1": {"id": 1, "name": "vertical", "op": {"type": "Cylinder", "radius": 10, "height": 40, "segments": 32}}, "2": {"id": 2, "name": "horizontal", "op": {"type": "Cylinder", "radius": 10, "height": 40, "segments": 32}}, "3": {"id": 3, "name": "h_rotated", "op": {"type": "Rotate", "child": 2, "angles": {"x": 0, "y": 90, "z": 0}}}, "4": {"id": 4, "name": "h_pos", "op": {"type": "Translate","child": 3, "offset": {"x": -20, "y": 0, "z": 20}}}, "5": {"id": 5, "name": "cross", "op": {"type": "Union", "left": 1, "right": 4}} }, "materials": {}, "part_materials": {}, "roots": [{"root": 5, "material": "default"}] }Expected volume:
2·π·10²·40 − (16/3)·10³ = 25132.74 − 5333.33 = 19799.41 mm³Actual volume: 24971.6 mm³ (kernel reports approximately 2× the cylinder volume — Steinmetz region is being double-counted)
Affected mecheval tasks
a3-cross-shaft-01— 23 deterministic failures all reporting actual volume ≈ 24971.6, 0/31 pass-rate across Opus / Sonnet / Haiku / GPT-5 / GPT-5-mini / GPT-4o-miniProbable root cause
In the SSI stage of the boolean pipeline, two perpendicular cylinders intersect along a pair of curves (two ellipse-like seams forming the boundary of the Steinmetz solid). If the SSI is missed or the cylindrical-face splits aren't classified, the trim/sew stage emits a result that contains both full cylinders' lateral surfaces overlapping in space — manifold is closed (each cylinder is closed on its own), but the volume integral over the unified mesh effectively double-counts the overlap region.
Worth a
cargo test -p vcad-kernel-booleans --features debug-boolean test_perp_cylinder_union -- --nocapturerun focused on theCylinder × CylinderSSI case incrates/vcad-kernel-booleans/src/ssi.rs.Acceptance
Volume of
cylinder(r=10, h=40, axis=Z) ∪ cylinder(r=10, h=40, axis=X, centered at z=20)returns19799.41 ± 1%.Refs