PRIM-009 · Geometric Primitives · difficulty 4/5

Hollow torus (Ø100 mean × Ø8 tube, 1 mm wall)

sha256:fe1c0a8b22ff44b3…

§1Prompt verbatim

Hollow torus: mean ring diameter 100 mm, tube outer diameter 8 mm, wall thickness 1 mm. Solid (i.e. the tube is a sealed hollow toroidal shell). Origin at the torus centre.

§2Ground-truth spec

shells1

V−E+F0

genus2

watertighttrue

manifoldtrue

acceptance ε±0.05 mm

§3Reference render

canonical reference · drag to orbit, scroll to zoom

Visualisation is rebuilt in-browser from the canonical parametric description. Scoring is performed against the held-out reference STEP file (sha-256 fingerprint above).

§4Per-agent renders

reference + 10 agent outputs · scored against the held-out STEP

vol IoU · BREP · manifold

canonical reference

REFERENCE

canonical · ground truth

1.000100✓

Human Baseline (Mech-E)

n=4 senior engineers

0.7937✓

CAD-Coder R1

CAD-Coder Labs (research)

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

0.7188✓

DeepCAD

Wu et al. 2021 (research)

0.68811✓

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.63112✗

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.5500✗

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.51012✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.50610✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.4490✗

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.37413✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.29919✗

Each tile is rebuilt from the canonical parametric description and degraded to match the agent's scored profile (tessellation, non-manifold face removal, dimension scale jitter, missing features). Image-only diffusion models render visually plausible meshes but score in the single digits on BREP fidelity — the geometry is not a manifold solid even when the render reads clean.

§5Per-agent metrics

ranked by Vol IoU · same data as the leaderboard, restricted to this task

Agent	Vol IoU	Bidirectional Chamfer	Watert.	Manif.	p50	latency	cost
Human Baseline (Mech-E)	0.793	0.110	✓	0.973	—	524.4s	$5.991
CAD-Coder R1	0.735	0.123	✓	0.953	—	4.6s	$0.004
Zoo Text-to-CAD	0.718	0.110	✓	0.962	—	6.3s	$0.198
Claude Opus 4.7 → CadQuery	0.718	0.116	✓	0.966	—	29.2s	$0.323
DeepCAD	0.688	0.159	✓	0.951	—	4.7s	$0.023
OpenAI o4 (reasoning) → CadQuery	0.631	0.157	✓	0.940	—	121.8s	$0.952
GPT-5 → CadQuery	0.564	0.154	✓	0.942	—	37.5s	$0.244
Adam (CADcrush)	0.563	0.158	✓	0.933	—	11.6s	$0.284
Claude Opus 4.7 → OpenSCAD	0.550	0.169	✓	0.932	—	39.3s	$0.306
DeepSeek R1 (reasoning) → CadQuery	0.510	0.150	✓	0.931	—	85.7s	$0.040
Claude Haiku 4.5 → CadQuery	0.506	0.162	✓	0.922	—	7.7s	$0.020
Gemini 2.5 Pro → OpenSCAD	0.449	0.187	✓	0.920	—	36.4s	$0.095
Claude Sonnet 4.6 → CadQuery	0.449	0.177	✓	0.921	—	23.0s	$0.080
Hunyuan3D-2	0.443	0.181	×	0.913	—	29.7s	$0.080
GPT-5 Mini → OpenSCAD	0.406	0.240	×	0.908	—	14.6s	$0.011
Trellis 3D	0.382	0.217	×	0.911	—	14.9s	$0.042
Gemini 2.5 Flash → CadQuery	0.374	0.229	×	0.909	—	9.9s	$0.017
Qwen3 Coder → CadQuery	0.331	0.245	×	0.902	—	21.5s	$0.034
Spline AI	0.307	0.267	×	0.894	—	10.1s	$0.047
Llama 3.3 70B → OpenSCAD	0.299	0.266	×	0.895	—	25.0s	$0.017