FUNC-003 · Functional Intent · FEA-Gated · difficulty 5/5

Heat-sink for 60 W CoB LED

sha256:fab02ec01dd30a11…

§1Prompt verbatim

Design a passive heat sink in 6063-T5 aluminium that holds a 60 W CoB LED below 85 °C junction temperature in still air at 30 °C ambient. LED footprint Ø 22 mm, mounting requires three M3 holes on a 28 mm PCD. Footprint envelope ≤ 80 × 80 mm, height ≤ 50 mm. Output STEP.

§2Ground-truth spec

shells1

watertighttrue

manifoldtrue

acceptance ε±0.1 mm

featuresLED_mount_M3_x3, fin_array

Functional intent task — scoring is FEA/CFD-gated, not geometric. A part with vol_iou ≈ 1 to the reference can still fail if the surface area is < 0.04 m².

§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

0.6749✗

OpenAI o4 (reasoning) → CadQuery

OpenAI + CadQuery 2.4

0.55811✗

Claude Opus 4.7 → CadQuery

Anthropic + CadQuery 2.4

0.49810✗

Claude Sonnet 4.6 → CadQuery

Anthropic + CadQuery 2.4

DeepSeek R1 (reasoning) → CadQuery

DeepSeek + CadQuery 2.4

0.36314✗

GPT-5 → CadQuery

OpenAI + CadQuery 2.4

0.35115✗

Claude Opus 4.7 → OpenSCAD

Anthropic + OpenSCAD 2024.06

0.3360✗

Qwen3 Coder → CadQuery

Alibaba + CadQuery 2.4

0.32120✗

Gemini 2.5 Flash → CadQuery

Google + CadQuery 2.4

Gemini 2.5 Pro → OpenSCAD

Google + OpenSCAD 2024.06

0.2120✗

Claude Haiku 4.5 → CadQuery

Anthropic + CadQuery 2.4

0.18132✗

DeepCAD

Wu et al. 2021 (research)

0.13137✗

CAD-Coder R1

CAD-Coder Labs (research)

0.10448✗

Llama 3.3 70B → OpenSCAD

Meta + OpenSCAD 2024.06

0.025104✗

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.	FeatRec	Min-Wall Compliance	FEA-Yield Pass	p50	latency	cost
Human Baseline (Mech-E)	✓	0.948	0.856	0.897	0.815	—	855.0s	$4.832
OpenAI o4 (reasoning) → CadQuery	✓	0.937	0.721	0.624	0.644	—	112.7s	$0.990
Claude Opus 4.7 → CadQuery	✓	0.929	0.680	0.631	0.534	—	30.4s	$0.351
Claude Sonnet 4.6 → CadQuery	✓	0.914	0.658	0.606	0.505	—	23.3s	$0.076
Zoo Text-to-CAD	✓	0.914	0.759	0.667	0.535	—	5.8s	$0.213
DeepSeek R1 (reasoning) → CadQuery	×	0.903	0.617	0.566	0.468	—	83.0s	$0.046
GPT-5 → CadQuery	×	0.903	0.714	0.651	0.477	—	47.0s	$0.212
Claude Opus 4.7 → OpenSCAD	×	0.904	0.522	0.605	0.410	—	22.5s	$0.265
Qwen3 Coder → CadQuery	×	0.896	0.622	0.520	0.391	—	18.9s	$0.029
Gemini 2.5 Flash → CadQuery	×	0.895	0.573	0.564	0.388	—	8.7s	$0.020
Adam (CADcrush)	×	0.884	0.657	0.619	0.408	—	8.5s	$0.257
Gemini 2.5 Pro → OpenSCAD	×	0.882	0.482	0.499	0.391	—	33.3s	$0.101
Claude Haiku 4.5 → CadQuery	×	0.875	0.505	0.434	0.261	—	8.3s	$0.020
DeepCAD	×	0.869	0.470	0.345	0.179	—	5.9s	$0.023
CAD-Coder R1	×	0.866	0.646	0.473	0.275	—	7.6s	$0.004
Llama 3.3 70B → OpenSCAD	×	0.854	0.388	0.420	0.217	—	19.8s	$0.016
GPT-5 Mini → OpenSCAD	×	0.853	0.398	0.469	0.218	—	13.4s	$0.011
Trellis 3D	×	0.850	0.212	0.197	0.093	—	13.7s	$0.049
Spline AI	×	0.850	0.086	0.099	0.047	—	9.8s	$0.038
Hunyuan3D-2	×	0.850	0.190	0.212	0.100	—	36.4s	$0.058