StepScore CAD Benchmark
44-metric benchmarking platform that scores AI-generated CAD models against human-calibrated quality thresholds.
What it does and how it connects
StepScore is a benchmarking platform that scores AI-generated CAD models against reference designs. A 44-metric geometry comparison engine loads two STEP files, aligns them with PCA and ICP, and computes metrics across 7 categories covering distance, volume, topology, curvature, occupancy, cross-section, and rendering.
The benchmark runs 200 jobs across 10 part families and 2 prompt difficulty levels, executed against Claude Sonnet 4, and the README reports a 99.5 percent success rate with an average quality score of 53.9 out of 100. Thresholds are human-calibrated from 199 manually labeled pass and fail pairs, reaching 88.9 percent accuracy and 98.6 percent precision.
The repository also ships a Streamlit dashboard, a CLI, a batch harness runner, a 3D viewer, and a threshold tuning pipeline, with versioned platform snapshots from the initial prototype through the MVP backend.
Alternatives
Compare side by side- P3D-Bench
Benchmarks multimodal LLMs on parametric 3D generation and assemblies from text or images, rather than scoring AI-generated STEP models against references with 44 geometry metrics.
- CodeToCAD
A code-first framework for defining CAD designs in Python, rather than a benchmark that scores AI-generated models.
Sources
Not verified
- License: no open-source license file; the GitHub API returns no license and the license endpoint returns 404, but the README states Proprietary. All rights reserved.
- No GitHub releases exist, so latest release fields are null.
- No independent user voice found: the repo has no public issues listed via the API.
- Pricing: none stated on the repo.
- FreeCAD as a connected tool: the directory record names FreeCAD; the README does not name it.
- Related Error Index entries: no matching entry verified.
Last verified November 6, 2026. Listings are informational. The evidence score measures public, verifiable evidence and is not an endorsement or a compatibility claim.