Deterministic 3D CAD CUA & B-Rep Geometric RLVR Infrastructure
Frontier AI cannot master physical engineering on 2D text or approximate meshes. We build the multimodal 3D CAD computer-use datasets and deterministic ISO-10303-21 B-Rep verification engines that unlock spatial intelligence for frontier models.
Why Verifiable 3D B-Rep Environments Matter
The four engineering principles guiding our 3D CAD CUA and geometric RLVR infrastructure.
Beyond 2D Text & Static Mesh Approximations
Physical manufacturing runs on watertight parametric B-Rep solids, not tessellated meshes or text descriptions. Frontier models need interactive 6-DoF CAD environments where they sketch, constrain, extrude, and verify exact solid geometry.
Deterministic B-Rep Truth Over VLM Judges
2D rendered screenshot judges hallucinate and miss internal bores or non-manifold edges. Our verify_step.py engine computes mathematical composite rewards R in [0, 1] directly from ISO-10303-21 .STEP boundary representations.
Multi-Suite Commercial & Open CAD Coverage
Real mechanical engineering spans diverse kernels. We capture verified human trajectories and native parametric deliverables across SolidWorks (51), PTC Creo (9), AutoCAD (1), FreeCAD (1), Autodesk Fusion (1), and Autodesk Inventor (1).
Zero-PII Curation & Industrial RLVR Scale
Every episode undergoes 9-pass CPU Vision OCR privacy scrubbing (72,070 raw frames audited down to 11,792 zero-PII PNGs), dynamic foreground window verification (82/82 bundles), and SHA-256 attestation across 401 public files.
From Mechanical CAD Foundry to Industrial B-Rep RLVR
How SpacetimeML built the canonical 3D CAD CUA dataset and deterministic .STEP reward engine.
The Mechanical CAD Expert Foundry
Built a specialized network of mechanical design engineers and manufacturing specialists recording synchronized high-resolution viewport keyframes, 237,579 desktop telemetry events, and native parametric feature trees.
Empirical Discovery of the 85–90% 3D CAD Failure Rate
Evaluated leading multimodal models on 4-view Third-Angle orthographic + isometric engineering drawings—proving that Claude 4.5 (15.0% Pass@1), GPT-5.2 (10.0%), and Gemini 3 (10.0%) fail 85–90% of 3D B-Rep reconstructions.
Deterministic ISO-10303-21 .STEP Verifier (verify_step.py)
Engineered our self-contained B-Rep geometric reward engine scoring Principal-Axis Bounding Box IoU (35%), Closed-Shell Volume/Area (25%), Euler-Poincaré Topology (20%), and Analytical Surface & Bore Radius Spectra (20%).
Public CAD-CUA Release (5.53 GB) & Frontier Lab RLVR Deployments
Released spacetime-ml/cad-cua (62 episodes, 48 ISO-128 blueprint pairs, 215 CAD deliverables including 98 .STEP solids, and the 20-task RLVR benchmark) under CC BY 4.0 while scaling private CAD-RLVR clusters for frontier AI labs.
The Verifiable Reward Paradigm for 3D Physical Engineering
Just as unit tests and compilers unlocked RLVR for code generation, deterministic ISO-10303-21 B-Rep verification (verify_step.py) provides the exact mathematical ground truth required for frontier models to master 3D mechanical CAD and spatial computer use.
Collaborate with our engineering team
We partner with frontier AI labs and industrial robotics teams to scale 3D CAD CUA and geometric RLVR post-training.