SEAL-001 · Sealing-Groove Design · difficulty 4/5

AS568-214 face-seal groove

sha256:44c01ef0b1199aa0…

§1Prompt verbatim

Design the female face-seal groove for an AS568-214 O-ring (Ø 0.984 in cross-section 0.139 in / 3.53 mm). Sealed pressure 10 MPa static, hydraulic. Apply standard squeeze (16-25 %) and groove-fill (75-85 %).

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.05 mm

featuresgroove_AS568-214

§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

0.8329✓

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

0.6898✓

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.6608✗

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.6479✓

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.5909✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.5450✗

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.50513✗

CAD-Coder R1

CAD-Coder Labs (research)

0.45812✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

0.38413✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.3590✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

0.28618✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.26917✗

DeepCAD

Wu et al. 2021 (research)

0.19726✗

Llama 3.3 70B → OpenSCAD

Meta + 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	Standards Compliance	P@1	p50	latency	cost
Human Baseline (Mech-E)	✓	0.969	0.114	0.947	1.000	—	658.5s	$5.118
Claude Opus 4.7 → CadQuery	✓	0.950	0.264	0.730	0.000	—	49.2s	$0.376
OpenAI o4 (reasoning) → CadQuery	✓	0.946	0.251	0.741	0.000	—	107.1s	$1.014
GPT-5 → CadQuery	✓	0.956	0.226	0.674	0.000	—	29.2s	$0.175
DeepSeek R1 (reasoning) → CadQuery	✓	0.939	0.317	0.533	0.000	—	120.5s	$0.043
Adam (CADcrush)	✓	0.937	0.160	0.711	0.000	—	9.5s	$0.301
Zoo Text-to-CAD	✓	0.945	0.159	0.696	0.000	—	6.7s	$0.180
Claude Sonnet 4.6 → CadQuery	✓	0.937	0.186	0.625	0.000	—	21.2s	$0.080
Claude Opus 4.7 → OpenSCAD	✓	0.928	0.255	0.618	0.000	—	23.3s	$0.308
Gemini 2.5 Flash → CadQuery	✓	0.929	0.287	0.569	0.000	—	15.1s	$0.018
CAD-Coder R1	✓	0.914	0.224	0.474	0.000	—	5.0s	$0.004
Qwen3 Coder → CadQuery	×	0.912	0.335	0.474	0.000	—	13.0s	$0.035
Gemini 2.5 Pro → OpenSCAD	×	0.908	0.281	0.508	0.000	—	30.1s	$0.073
GPT-5 Mini → OpenSCAD	×	0.896	0.295	0.374	0.000	—	12.1s	$0.010
Claude Haiku 4.5 → CadQuery	×	0.893	0.304	0.457	0.000	—	6.9s	$0.016
DeepCAD	×	0.881	0.291	0.265	0.000	—	4.7s	$0.022
Llama 3.3 70B → OpenSCAD	×	0.877	0.396	0.340	0.000	—	25.3s	$0.023
Trellis 3D	×	0.858	0.542	0.047	0.000	—	8.5s	$0.054
Hunyuan3D-2	×	0.851	0.503	0.045	0.000	—	32.5s	$0.059
Spline AI	×	0.850	0.586	0.020	0.000	—	5.9s	$0.043