BREP-002 · BREP Fidelity · difficulty 4/5

G1-only loft (tangent discontinuity)

sha256:0f3ac01eb5deaa11…

§1Prompt verbatim

Loft three closed sketches: circle Ø 30 at z=0, square 30 × 30 at z=20, circle Ø 30 at z=40. Match tangents (G1) but NOT curvature (G2). Output STEP. Reference uses an unguided ruled loft.

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.05 mm

featuresg1_loft, no_g2_continuity

§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✓

Zoo Text-to-CAD

Zoo (KittyCAD)

0.6977✓

Human Baseline (Mech-E)

n=4 senior engineers

0.6918✓

CAD-Coder R1

CAD-Coder Labs (research)

0.6787✓

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

0.55913✗

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.52610✗

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.51310✗

DeepCAD

Wu et al. 2021 (research)

0.42214✗

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

0.40213✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.23824✗

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.0760✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

GPT-5 Mini → OpenSCAD

OpenAI + OpenSCAD 2024.06

0.000100✗

no manifold solid produced

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

—59✗

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.000101✗

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.	Euler-Poincaré Compliance	STEP RT	FeatRec	p50	latency	cost
Zoo Text-to-CAD	✓	0.956	✓	0.069	0.768	—	5.2s	$0.175
Human Baseline (Mech-E)	✓	0.957	✓	0.076	0.922	—	657.7s	$7.025
CAD-Coder R1	✓	0.955	✓	0.071	0.670	—	5.0s	$0.004
DeepSeek R1 (reasoning) → CadQuery	✓	0.946	✓	0.095	0.627	—	86.9s	$0.044
Adam (CADcrush)	✓	0.945	✓	0.111	0.730	—	10.3s	$0.317
Claude Opus 4.7 → CadQuery	✓	0.930	✓	0.127	0.704	—	33.9s	$0.347
OpenAI o4 (reasoning) → CadQuery	✓	0.925	×	0.101	0.792	—	118.3s	$1.277
GPT-5 → CadQuery	✓	0.926	×	0.096	0.680	—	31.6s	$0.226
DeepCAD	✓	0.919	×	0.136	0.458	—	4.3s	$0.024
Claude Sonnet 4.6 → CadQuery	×	0.906	×	0.130	0.719	—	21.0s	$0.057
Claude Haiku 4.5 → CadQuery	×	0.886	×	0.238	0.542	—	6.8s	$0.021
Gemini 2.5 Pro → OpenSCAD	×	0.862	×	—	0.481	—	21.1s	$0.107
Claude Opus 4.7 → OpenSCAD	×	0.850	×	—	0.519	—	26.2s	$0.336
Trellis 3D	×	0.850	×	—	0.200	—	13.8s	$0.049
Spline AI	×	0.850	×	—	0.089	—	10.2s	$0.040
GPT-5 Mini → OpenSCAD	×	0.850	×	1.005	0.410	—	16.0s	$0.012
Gemini 2.5 Flash → CadQuery kernel error: BRepCheck_NotClosed	×	0.000	—	—	—	—	11.5s	$0.019
Llama 3.3 70B → OpenSCAD	×	0.850	×	1.011	0.447	—	17.3s	$0.020
Qwen3 Coder → CadQuery kernel error: BRepCheck_NotClosed	×	0.000	—	—	—	—	17.3s	$0.031
Hunyuan3D-2	×	0.850	×	1.046	0.208	—	34.1s	$0.067