MECH-005 · Parametric Mechanical Parts · difficulty 3/5

Cam-follower lever (eccentric pivot)

sha256:fe22ac01bdc40e90…

§1Prompt verbatim

Lever 110 mm long: pivot bore Ø 8 H7 at one end, follower roller pin bore Ø 6 H7 at the other end. Eccentric pivot at 32 mm from the load end, lever thickness 6 mm, web fillet R 4 mm on both faces. Two outer profile cusps R 8 mm. Output STEP.

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.05 mm

featuresbore_H7_8, bore_H7_6, fillet_R4

§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

0.8468✓

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.6997✓

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.6768✓

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

Wu et al. 2021 (research)

0.54712✗

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.53410✗

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.53111✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.52112✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

0.49012✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.4510✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.4050✗

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.38816✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

no manifold solid produced

Spline AI

Spline.design

—2✗

no manifold solid produced

CAD-Coder R1

CAD-Coder Labs (research)

—74✗

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	Watert.	Manif.	Named-Dimension RMSE	GD&T Compliance	FeatRec	P@1	p50	latency	cost
Human Baseline (Mech-E)	✓	0.971	0.105	0.885	0.916	1.000	—	618.6s	$5.045
OpenAI o4 (reasoning) → CadQuery	✓	0.956	0.247	0.662	0.719	0.000	—	101.5s	$1.196
Claude Sonnet 4.6 → CadQuery	✓	0.951	0.276	0.534	0.649	0.000	—	22.8s	$0.063
GPT-5 → CadQuery	✓	0.957	0.191	0.583	0.736	0.000	—	41.0s	$0.227
Zoo Text-to-CAD	✓	0.951	0.151	0.678	0.803	0.000	—	6.9s	$0.159
Claude Opus 4.7 → CadQuery	✓	0.946	0.225	0.633	0.704	0.000	—	40.3s	$0.403
Adam (CADcrush)	✓	0.943	0.184	0.605	0.642	0.000	—	10.5s	$0.252
DeepCAD	✓	0.927	0.340	0.389	0.467	0.000	—	5.8s	$0.021
Gemini 2.5 Flash → CadQuery	✓	0.931	0.311	0.429	0.551	0.000	—	14.7s	$0.022
DeepSeek R1 (reasoning) → CadQuery	✓	0.935	0.282	0.518	0.620	0.000	—	75.6s	$0.044
Claude Haiku 4.5 → CadQuery	✓	0.929	0.307	0.380	0.516	0.000	—	7.4s	$0.022
Qwen3 Coder → CadQuery	✓	0.919	0.326	0.421	0.624	0.000	—	19.0s	$0.036
Gemini 2.5 Pro → OpenSCAD	✓	0.917	0.254	0.424	0.545	0.000	—	27.4s	$0.103
Claude Opus 4.7 → OpenSCAD	×	0.912	0.264	0.461	0.532	0.000	—	30.3s	$0.363
Llama 3.3 70B → OpenSCAD	×	0.905	0.345	0.289	0.439	0.000	—	17.1s	$0.022
GPT-5 Mini → OpenSCAD	×	0.882	0.276	0.294	0.429	0.000	—	13.5s	$0.009
Trellis 3D	×	0.882	0.547	0.055	0.198	0.000	—	12.1s	$0.057
Hunyuan3D-2	×	0.875	0.464	0.055	0.198	0.000	—	29.6s	$0.082
Spline AI kernel error: BRepCheck_NotClosed	×	0.000	—	—	—	0.000	—	8.7s	$0.043
CAD-Coder R1 kernel error: BRepCheck_NotClosed	×	0.000	—	—	—	0.000	—	5.5s	$0.005