PARAM-013 · Constraint Solving & Editability · difficulty 4/5

Editable bracket (length+30 %, hole→M8)

sha256:fab07d2c5e914421…

§1Prompt verbatim

Build the L-bracket from MECH-014, then perform two parametric edits in sequence: (1) increase the long leg from 60 → 78 mm; (2) change the through-hole from M6 clearance to M8 clearance (Ø 9.0 mm). Topology must remain valid throughout.

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.1 mm

parametric edits

leg_long : 60 → 78 (ΔV expected 3600 mm³)

hole_d : 6.6 → 9 (ΔV expected -147 mm³)

§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✓

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.75110✓

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.72810✓

Human Baseline (Mech-E)

n=4 senior engineers

0.71810✓

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.64812✗

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

0.53411✗

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.52610✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.4590✗

CAD-Coder R1

CAD-Coder Labs (research)

0.42915✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.36516✗

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

0.32615✗

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.22525✗

DeepCAD

Wu et al. 2021 (research)

0.16732✗

no manifold solid produced

Claude Opus 4.7 → OpenSCAD

Anthropic + 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.	ParamEdit	ConSolve	Parametric Range Integrity	P@1	p50	latency	cost
GPT-5 → CadQuery	✓	0.958	0.693	0.713	0.726	1.000	—	39.1s	$0.226
OpenAI o4 (reasoning) → CadQuery	✓	0.954	0.866	0.792	0.760	0.000	—	81.5s	$1.086
Human Baseline (Mech-E)	✓	0.957	0.853	0.801	0.845	0.000	—	711.6s	$5.129
DeepSeek R1 (reasoning) → CadQuery	✓	0.945	0.690	0.730	0.648	0.000	—	122.3s	$0.037
Claude Opus 4.7 → CadQuery	✓	0.951	0.732	0.746	0.756	0.000	—	46.9s	$0.389
Adam (CADcrush)	✓	0.952	0.740	0.748	0.681	0.000	—	10.1s	$0.237
Gemini 2.5 Flash → CadQuery	✓	0.939	0.626	0.562	0.515	0.000	—	9.5s	$0.017
Zoo Text-to-CAD	✓	0.937	0.661	0.734	0.651	0.000	—	7.7s	$0.192
Qwen3 Coder → CadQuery	✓	0.930	0.647	0.660	0.604	0.000	—	12.7s	$0.035
Claude Sonnet 4.6 → CadQuery	✓	0.936	0.737	0.807	0.651	0.000	—	16.2s	$0.065
Gemini 2.5 Pro → OpenSCAD	✓	0.923	0.599	0.585	0.544	0.000	—	26.4s	$0.090
CAD-Coder R1	✓	0.914	0.615	0.610	0.549	0.000	—	5.0s	$0.005
Claude Haiku 4.5 → CadQuery	×	0.908	0.547	0.486	0.399	0.000	—	10.2s	$0.017
GPT-5 Mini → OpenSCAD	×	0.904	0.449	0.426	0.424	0.000	—	17.3s	$0.010
Llama 3.3 70B → OpenSCAD	×	0.883	0.411	0.444	0.345	0.000	—	24.0s	$0.021
DeepCAD	×	0.875	0.297	0.320	0.245	0.000	—	4.4s	$0.019
Claude Opus 4.7 → OpenSCAD kernel error: BRepCheck_NotClosed	×	0.000	—	—	—	0.000	—	25.8s	$0.316
Trellis 3D	×	0.850	0.057	0.056	0.036	0.000	—	11.7s	$0.050
Spline AI	×	0.850	0.047	0.046	0.018	0.000	—	7.0s	$0.043
Hunyuan3D-2	×	0.850	0.048	0.046	0.037	0.000	—	25.2s	$0.076