PRIM-002 · Geometric Primitives · difficulty 1/5

Sphere with planar cap

sha256:12bf4e0c44a1aa01…

§1Prompt verbatim

Solid sphere of radius 25 mm cut by the plane z = 18 mm; keep the −z portion. Origin at sphere centre. Output a watertight solid.

§2Ground-truth spec

bbox50.0 × 50.0 × 43.0 mm

shells1

V−E+F2

genus0

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✓

CAD-Coder Labs (research)

0.83810✓

DeepCAD

Wu et al. 2021 (research)

0.7788✓

Human Baseline (Mech-E)

n=4 senior engineers

0.7758✓

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.7617✓

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.7417✓

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

0.6598✓

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.62910✓

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.5930✗

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.56912✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

0.54912✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.5370✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

0.52012✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

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.	P@1	p50	latency	cost
Zoo Text-to-CAD	0.884	0.135	✓	0.975	1.000	—	7.3s	$0.150
CAD-Coder R1	0.838	0.136	✓	0.971	0.000	—	5.9s	$0.006
DeepCAD	0.778	0.100	✓	0.964	0.000	—	5.1s	$0.020
Human Baseline (Mech-E)	0.775	0.129	✓	0.966	0.000	—	697.9s	$6.817
Claude Sonnet 4.6 → CadQuery	0.761	0.148	✓	0.959	0.000	—	20.8s	$0.065
OpenAI o4 (reasoning) → CadQuery	0.741	0.130	✓	0.959	0.000	—	116.0s	$1.134
GPT-5 → CadQuery	0.696	0.158	✓	0.953	0.000	—	35.9s	$0.237
Adam (CADcrush)	0.663	0.139	✓	0.959	0.000	—	8.2s	$0.219
Claude Opus 4.7 → CadQuery	0.659	0.153	✓	0.955	0.000	—	42.1s	$0.347
Gemini 2.5 Flash → CadQuery	0.629	0.132	✓	0.951	0.000	—	12.6s	$0.023
Llama 3.3 70B → OpenSCAD	0.605	0.138	✓	0.941	0.000	—	15.9s	$0.024
Hunyuan3D-2	0.598	0.139	✓	0.936	0.000	—	44.0s	$0.078
Gemini 2.5 Pro → OpenSCAD	0.593	0.141	✓	0.941	0.000	—	29.6s	$0.088
DeepSeek R1 (reasoning) → CadQuery	0.569	0.172	✓	0.941	0.000	—	73.3s	$0.048
Qwen3 Coder → CadQuery	0.549	0.140	✓	0.936	0.000	—	13.0s	$0.025
Claude Opus 4.7 → OpenSCAD	0.537	0.185	✓	0.931	0.000	—	38.5s	$0.271
GPT-5 Mini → OpenSCAD	0.520	0.190	✓	0.929	0.000	—	10.2s	$0.008
Claude Haiku 4.5 → CadQuery	0.477	0.174	✓	0.925	0.000	—	9.2s	$0.018
Spline AI	0.416	0.207	✓	0.917	0.000	—	8.1s	$0.042
Trellis 3D	0.403	0.190	✓	0.913	0.000	—	11.9s	$0.054