MECH-002 · Parametric Mechanical Parts · difficulty 4/5

Bearing block — two 6204 deep-groove bearings

sha256:a4f001bedc20cf91…

§1Prompt verbatim

Bearing block to house two SKF 6204 bearings (Ø 47 OD, Ø 20 ID, 14 mm wide) on a common axis 60 mm apart on centres. Block envelope 80 × 50 × 40 mm. Two M5 mounting bolts on a 70 × 30 mm rectangle, through-hole with 8 × 1 mm counter-bores. Bearing seats Ø 47 H7 with 0.5 × 45° lead-in chamfer.

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.05 mm

featuresbearing_seat_47H7_x2, M5_clearance_x2, counterbore_8x1

§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✓

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.66112✗

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.60512✗

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.5889✗

Human Baseline (Mech-E)

n=4 senior engineers

0.5828✗

CAD-Coder R1

CAD-Coder Labs (research)

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.52811✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.4640✗

DeepCAD

Wu et al. 2021 (research)

0.46111✗

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.41412✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

0.38113✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.3620✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

0.31116✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

no manifold solid produced

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

—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	p50	latency	cost
GPT-5 → CadQuery	✓	0.945	0.281	0.577	0.671	—	53.3s	$0.242
Claude Opus 4.7 → CadQuery	✓	0.938	0.236	0.659	0.710	—	26.8s	$0.307
Zoo Text-to-CAD	✓	0.940	0.151	0.635	0.691	—	4.6s	$0.146
Claude Sonnet 4.6 → CadQuery	✓	0.939	0.241	0.598	0.721	—	20.3s	$0.063
DeepSeek R1 (reasoning) → CadQuery	✓	0.933	0.271	0.537	0.643	—	105.1s	$0.039
Human Baseline (Mech-E)	✓	0.938	0.077	0.812	0.955	—	805.7s	$4.846
CAD-Coder R1	✓	0.938	0.320	0.479	0.652	—	5.0s	$0.005
Adam (CADcrush)	✓	0.938	0.175	0.678	0.631	—	9.8s	$0.310
Gemini 2.5 Flash → CadQuery	✓	0.926	0.293	0.444	0.537	—	15.1s	$0.021
Claude Opus 4.7 → OpenSCAD	✓	0.917	0.261	0.463	0.593	—	32.0s	$0.330
DeepCAD	✓	0.915	0.319	0.347	0.488	—	4.9s	$0.024
Llama 3.3 70B → OpenSCAD	✓	0.915	0.367	0.294	0.445	—	20.5s	$0.019
Qwen3 Coder → CadQuery	×	0.909	0.308	0.433	0.595	—	18.5s	$0.033
Gemini 2.5 Pro → OpenSCAD	×	0.906	0.301	0.447	0.512	—	33.8s	$0.095
GPT-5 Mini → OpenSCAD	×	0.898	0.323	0.297	0.423	—	16.4s	$0.010
Claude Haiku 4.5 → CadQuery	×	0.900	0.321	0.388	0.471	—	8.2s	$0.024
Hunyuan3D-2	×	0.863	0.512	0.052	0.203	—	37.4s	$0.070
Trellis 3D	×	0.851	0.501	0.058	0.214	—	9.9s	$0.054
Spline AI	×	0.850	0.579	0.027	0.087	—	6.5s	$0.045
OpenAI o4 (reasoning) → CadQuery kernel error: BRepCheck_NotClosed	×	0.000	—	—	—	—	139.1s	$1.130