DFMCNC-002 · DFM · 3-Axis CNC · difficulty 4/5

3-ax-machinable manifold block

sha256:ab02f10c8e51dd23…

§1Prompt verbatim

Solid 80 × 60 × 40 mm aluminium block. Three Ø 6 H7 bores from the top, two Ø 4 H7 bores from each side, all intersecting an internal Ø 8 manifold channel running along the long axis. The geometry must be fully machinable on a 3-axis VMC with Ø 6 → Ø 4 → Ø 1 tools, plus a 90° drill-rotation between sides.

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.05 mm

featuresmanifold_channel, bore_H7_x7

§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)

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.5869✗

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

0.57010✗

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.56811✗

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.54110✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.48011✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.4460✗

CAD-Coder R1

CAD-Coder Labs (research)

0.43215✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.4090✗

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.37616✗

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.34716✗

DeepCAD

Wu et al. 2021 (research)

0.33418✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.31919✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

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.	DFM	Min-Wall Compliance	CAM Reachability	P@1	p50	latency	cost
Human Baseline (Mech-E)	✓	0.967	92.6	0.885	0.795	1.000	—	901.4s	$6.141
Zoo Text-to-CAD	✓	0.936	81.8	0.623	0.706	0.000	—	7.1s	$0.153
Gemini 2.5 Flash → CadQuery	✓	0.937	75.2	0.569	0.585	0.000	—	15.2s	$0.019
Claude Opus 4.7 → CadQuery	✓	0.933	80.8	0.705	0.644	0.000	—	29.2s	$0.316
GPT-5 → CadQuery	✓	0.933	78.6	0.615	0.596	0.000	—	49.4s	$0.175
Claude Sonnet 4.6 → CadQuery	✓	0.928	79.6	0.631	0.608	0.000	—	21.4s	$0.070
Qwen3 Coder → CadQuery	✓	0.924	75.7	0.503	0.483	0.000	—	15.2s	$0.032
Adam (CADcrush)	✓	0.925	82.1	0.638	0.649	0.000	—	11.4s	$0.312
OpenAI o4 (reasoning) → CadQuery	✓	0.919	86.7	0.676	0.652	0.000	—	97.4s	$1.101
Claude Opus 4.7 → OpenSCAD	✓	0.917	75.6	0.563	0.475	0.000	—	33.2s	$0.253
CAD-Coder R1	×	0.911	77.4	0.538	0.543	0.000	—	4.5s	$0.006
Gemini 2.5 Pro → OpenSCAD	✓	0.916	74.6	0.494	0.486	0.000	—	33.8s	$0.102
DeepSeek R1 (reasoning) → CadQuery	×	0.906	77.5	0.588	0.582	0.000	—	112.7s	$0.034
Llama 3.3 70B → OpenSCAD	×	0.906	65.9	0.408	0.384	0.000	—	19.0s	$0.016
DeepCAD	×	0.903	63.1	0.385	0.363	0.000	—	6.1s	$0.024
Claude Haiku 4.5 → CadQuery	×	0.899	68.1	0.460	0.414	0.000	—	7.7s	$0.024
GPT-5 Mini → OpenSCAD	×	0.893	71.8	0.430	0.367	0.000	—	14.5s	$0.009
Spline AI	×	0.860	48.8	0.099	0.049	0.000	—	5.6s	$0.033
Hunyuan3D-2	×	0.853	50.8	0.238	0.092	0.000	—	30.7s	$0.059
Trellis 3D	×	0.852	52.3	0.212	0.086	0.000	—	10.8s	$0.043