SHEET-003 · Sheet-Metal Bodies · difficulty 3/5

U-channel bracket, 1.5 mm Al, k=0.40

sha256:311fac0bb472ea91…

§1Prompt verbatim

Sheet-metal U-channel from 1.5 mm Al-5052: outside dimensions 80 × 30 × 25 mm tall, bend radius 1.5 mm (inside), k-factor 0.40. Two M5 clearance holes (Ø 5.5) on the base, centred 15 mm from each end. Output the folded body and report the unfolded flat pattern dimensions.

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.1 mm

featuresbend_x2, M5_clearance_x2

§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

0.8486✓

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

0.77011✓

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.6840✗

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.62610✗

CAD-Coder R1

CAD-Coder Labs (research)

0.62010✗

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.5599✗

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.5579✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.49414✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

0.49213✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.3790✗

DeepCAD

Wu et al. 2021 (research)

0.37513✗

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.33216✗

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.	Named-Dimension RMSE	FeatRec	Wall-Thickness Uniformity	P@1	p50	latency	cost
Human Baseline (Mech-E)	✓	0.979	0.112	0.824	0.802	1.000	—	666.8s	$6.217
Claude Opus 4.7 → CadQuery	✓	0.960	0.193	0.787	0.649	1.000	—	30.4s	$0.382
Claude Opus 4.7 → OpenSCAD	✓	0.947	0.265	0.537	0.586	0.000	—	25.4s	$0.299
Claude Sonnet 4.6 → CadQuery	✓	0.958	0.200	0.739	0.650	0.000	—	22.4s	$0.065
Adam (CADcrush)	✓	0.942	0.164	0.640	0.649	0.000	—	10.5s	$0.257
OpenAI o4 (reasoning) → CadQuery	✓	0.942	0.248	0.722	0.653	0.000	—	106.8s	$1.028
CAD-Coder R1	✓	0.948	0.266	0.673	0.496	0.000	—	7.5s	$0.005
DeepSeek R1 (reasoning) → CadQuery	✓	0.939	0.257	0.684	0.582	0.000	—	74.1s	$0.039
Zoo Text-to-CAD	✓	0.942	0.141	0.730	0.706	0.000	—	7.6s	$0.201
Gemini 2.5 Flash → CadQuery	✓	0.937	0.302	0.580	0.526	0.000	—	11.6s	$0.021
GPT-5 → CadQuery	✓	0.938	0.253	0.652	0.611	0.000	—	39.3s	$0.239
Claude Haiku 4.5 → CadQuery	✓	0.925	0.355	0.541	0.433	0.000	—	9.0s	$0.016
Qwen3 Coder → CadQuery	✓	0.925	0.268	0.594	0.476	0.000	—	14.2s	$0.033
Gemini 2.5 Pro → OpenSCAD	×	0.912	0.266	0.523	0.474	0.000	—	35.7s	$0.077
DeepCAD	×	0.908	0.310	0.491	0.369	0.000	—	3.5s	$0.024
Llama 3.3 70B → OpenSCAD	×	0.899	0.381	0.442	0.402	0.000	—	16.7s	$0.022
GPT-5 Mini → OpenSCAD	×	0.896	0.351	0.386	0.407	0.000	—	10.8s	$0.009
Hunyuan3D-2	×	0.868	0.485	0.190	0.124	0.000	—	37.4s	$0.059
Trellis 3D	×	0.865	0.556	0.204	0.131	0.000	—	11.8s	$0.041
Spline AI	×	0.850	0.541	0.092	0.066	0.000	—	8.5s	$0.039