PRIM-007 · Geometric Primitives · difficulty 2/5

Right hexagonal prism with pitch fillet

sha256:4f8d1f2cba83a911…

§1Prompt verbatim

Hexagonal prism, across-flats 24 mm, height 12 mm, top edge filleted at R 0.4 mm. Solid, manifold, origin centred.

§2Ground-truth spec

bbox27.7 × 24.0 × 12.0 mm

shells1

V−E+F2

genus0

watertighttrue

manifoldtrue

acceptance ε±0.02 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)

OpenAI + CadQuery 2.4

0.7179✓

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.71712✓

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

CAD-Coder Labs (research)

0.69111✓

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.68810✓

DeepCAD

Wu et al. 2021 (research)

0.64311✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

0.61513✗

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.5269✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.5210✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

0.50610✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

no manifold solid produced

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

—50✗

no manifold solid produced

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

—61✗

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
Human Baseline (Mech-E)	0.945	0.105	✓	0.984	1.000	—	591.1s	$6.804
Zoo Text-to-CAD	0.815	0.111	✓	0.967	0.000	—	4.9s	$0.152
GPT-5 → CadQuery	0.717	0.139	✓	0.961	0.000	—	44.4s	$0.231
Claude Sonnet 4.6 → CadQuery	0.717	0.128	✓	0.950	0.000	—	22.2s	$0.060
Gemini 2.5 Flash → CadQuery	0.712	0.115	✓	0.950	0.000	—	13.3s	$0.021
Adam (CADcrush)	0.694	0.144	✓	0.953	0.000	—	11.9s	$0.229
CAD-Coder R1	0.691	0.152	✓	0.959	0.000	—	6.8s	$0.004
OpenAI o4 (reasoning) → CadQuery	0.688	0.157	✓	0.952	0.000	—	139.5s	$1.139
DeepCAD	0.643	0.140	✓	0.942	0.000	—	5.4s	$0.019
Qwen3 Coder → CadQuery	0.615	0.156	✓	0.937	0.000	—	17.9s	$0.033
Claude Opus 4.7 → CadQuery	0.613	0.172	✓	0.937	0.000	—	47.0s	$0.370
Hunyuan3D-2	0.593	0.160	✓	0.933	0.000	—	40.5s	$0.058
Llama 3.3 70B → OpenSCAD	0.526	0.178	✓	0.932	0.000	—	24.2s	$0.016
Gemini 2.5 Pro → OpenSCAD	0.521	0.154	✓	0.935	0.000	—	31.5s	$0.092
GPT-5 Mini → OpenSCAD	0.506	0.163	✓	0.923	0.000	—	14.1s	$0.008
Claude Opus 4.7 → OpenSCAD	0.485	0.150	✓	0.928	0.000	—	24.9s	$0.332
Trellis 3D	0.393	0.231	×	0.910	0.000	—	8.4s	$0.043
Spline AI	0.273	0.285	×	0.895	0.000	—	5.8s	$0.038
Claude Haiku 4.5 → CadQuery kernel error: BRepCheck_NotClosed	0.000	—	×	0.000	0.000	—	7.6s	$0.019
DeepSeek R1 (reasoning) → CadQuery kernel error: BRepCheck_NotClosed	0.000	—	×	0.000	0.000	—	81.4s	$0.036