KIN-008 · Kinematic Mechanisms · difficulty 5/5

Planetary gearset full mesh (sun + 3 planets + ring)

sha256:44ec01ef0bc20e11…

§1Prompt verbatim

Planetary gear stack: module 1.5, sun 21T, planets 21T (×3), ring 63T, all 20° pressure angle, face width 8 mm, ISO 53 profile. Carrier plate Ø 110 × 4 mm. The full stack must mesh — no tooth interference, planet teeth align with sun and ring simultaneously. Output assembly STEP.

§2Ground-truth spec

shells5

watertighttrue

manifoldtrue

acceptance ε±0.02 mm

featuressun_21T, planet_21T_x3, ring_63T, carrier

§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)

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.52312✗

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.46312✗

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.43411✗

CAD-Coder R1

CAD-Coder Labs (research)

0.37616✗

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

0.34214✗

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.33015✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.3150✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.2660✗

DeepCAD

Wu et al. 2021 (research)

0.26620✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.24723✗

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.20725✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

no manifold solid produced

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

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.	FeatRec	Mating Clearance	ParamEdit	P@1	p50	latency	cost
Human Baseline (Mech-E)	✓	0.967	0.927	0.926	0.849	1.000	—	591.4s	$6.072
Zoo Text-to-CAD	✓	0.943	0.756	0.639	0.703	0.000	—	5.1s	$0.179
Claude Opus 4.7 → CadQuery	✓	0.934	0.683	0.671	0.757	0.000	—	43.4s	$0.327
Adam (CADcrush)	✓	0.931	0.657	0.608	0.739	0.000	—	10.1s	$0.223
OpenAI o4 (reasoning) → CadQuery	✓	0.930	0.809	0.730	0.828	0.000	—	92.0s	$1.306
DeepSeek R1 (reasoning) → CadQuery	✓	0.916	0.680	0.567	0.740	0.000	—	83.7s	$0.034
GPT-5 → CadQuery	✓	0.917	0.678	0.657	0.696	0.000	—	31.9s	$0.232
CAD-Coder R1	×	0.911	0.616	0.541	0.599	0.000	—	6.0s	$0.005
Gemini 2.5 Flash → CadQuery	×	0.906	0.538	0.542	0.655	0.000	—	10.7s	$0.020
Claude Sonnet 4.6 → CadQuery	×	0.901	0.692	0.647	0.737	0.000	—	13.3s	$0.068
Gemini 2.5 Pro → OpenSCAD	×	0.898	0.547	0.453	0.548	0.000	—	32.4s	$0.107
Claude Opus 4.7 → OpenSCAD	×	0.890	0.535	0.514	0.638	0.000	—	26.6s	$0.342
DeepCAD	×	0.888	0.455	0.371	0.317	0.000	—	5.9s	$0.022
Claude Haiku 4.5 → CadQuery	×	0.885	0.522	0.440	0.508	0.000	—	10.2s	$0.023
Llama 3.3 70B → OpenSCAD	×	0.882	0.411	0.334	0.405	0.000	—	17.8s	$0.017
GPT-5 Mini → OpenSCAD	×	0.876	0.410	0.322	0.480	0.000	—	14.9s	$0.009
Spline AI	×	0.852	0.099	0.044	0.049	0.000	—	6.5s	$0.038
Trellis 3D	×	0.850	0.195	0.092	0.055	0.000	—	11.9s	$0.044
Qwen3 Coder → CadQuery kernel error: BRepCheck_NotClosed	×	0.000	—	—	—	0.000	—	13.6s	$0.026
Hunyuan3D-2	×	0.850	0.196	0.099	0.047	0.000	—	44.2s	$0.068