REVENG-005 · Reverse Engineering · difficulty 4/5

ABC dataset stepped pulley (multi-view)

sha256:c0a1b3d40fe201bc…

§1Prompt verbatim

From the supplied front/top/side ortho drawing (1:1, 600 dpi) reproduce the stepped V-belt pulley. Three steps, OD 80 / 60 / 40 mm, each 12 mm wide; belt grooves 38° included angle, depth 10 mm. Central Ø 16 H7 bore. All dimensions on the drawing are authoritative.

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.1 mm

featuresstep_pulley_x3, v_groove_x3, bore_H7_16

§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✓

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.46711✗

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.46610✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.4620✗

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.43212✗

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.37514✗

CAD-Coder R1

CAD-Coder Labs (research)

0.34616✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.3310✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.28017✗

DeepCAD

Wu et al. 2021 (research)

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.23124✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

no manifold solid produced

Human Baseline (Mech-E)

n=4 senior engineers

—96✗

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	Watert.	Manif.	Named-Dimension RMSE	FeatRec	p50	latency	cost
OpenAI o4 (reasoning) → CadQuery	0.635	✓	0.953	0.159	0.717	—	133.7s	$1.154
Hunyuan3D-2	0.610	✓	0.944	0.527	0.210	—	30.6s	$0.082
Claude Opus 4.7 → CadQuery	0.545	✓	0.931	0.208	0.745	—	26.9s	$0.334
Zoo Text-to-CAD	0.533	✓	0.926	0.146	0.703	—	5.2s	$0.147
Qwen3 Coder → CadQuery	0.491	✓	0.922	0.244	0.616	—	21.3s	$0.034
Trellis 3D	0.480	✓	0.925	0.482	0.199	—	11.2s	$0.051
DeepSeek R1 (reasoning) → CadQuery	0.467	✓	0.918	0.269	0.646	—	74.8s	$0.047
GPT-5 → CadQuery	0.466	✓	0.926	0.254	0.652	—	46.1s	$0.198
Gemini 2.5 Pro → OpenSCAD	0.462	✓	0.915	0.256	0.527	—	35.9s	$0.101
Claude Sonnet 4.6 → CadQuery	0.432	✓	0.916	0.223	0.691	—	13.5s	$0.067
Gemini 2.5 Flash → CadQuery	0.375	×	0.903	0.285	0.538	—	9.2s	$0.017
CAD-Coder R1	0.346	×	0.906	0.285	0.623	—	6.3s	$0.005
Claude Opus 4.7 → OpenSCAD	0.331	×	0.902	0.269	0.531	—	37.4s	$0.327
Claude Haiku 4.5 → CadQuery	0.280	×	0.893	0.340	0.513	—	9.7s	$0.016
DeepCAD	0.251	×	0.887	0.333	0.477	—	4.1s	$0.018
Adam (CADcrush)	0.232	×	0.885	0.167	0.656	—	11.1s	$0.285
Llama 3.3 70B → OpenSCAD	0.231	×	0.886	0.319	0.398	—	18.7s	$0.017
GPT-5 Mini → OpenSCAD	0.210	×	0.881	0.286	0.393	—	11.1s	$0.010
Spline AI	0.186	×	0.877	0.544	0.089	—	9.8s	$0.042
Human Baseline (Mech-E) kernel error: BRepCheck_NotClosed	0.000	×	0.000	—	—	—	683.5s	$5.262