3D Spatial & CAD Post-Training for Frontier AI Labs
Resolve the 85–90% frontier model failure rate on 3D mechanical engineering tasks. Partner with SpacetimeML for private multi-suite CAD CUA trajectories, held-out ISO-128 blueprint benchmarks, and deterministic ISO-10303-21 .STEP RLVR environments.
Multi-Suite 3D CAD CUA Trajectories
Expand beyond the public 62-episode release (5.53 GB) with custom enterprise trajectories across SolidWorks, PTC Creo, AutoCAD, FreeCAD, Autodesk Fusion, and Autodesk Inventor.
Deterministic .STEP B-Rep Reward Engine (verify_step.py)
Score candidate 3D models with exact composite reward R in [0, 1] across Principal-Axis BBox IoU (35%), Closed-Shell Volume/Area (25%), Euler-Poincaré Topology (20%), and Bore Radius Spectrum (20%).
Held-Out ISO-128 Blueprint RLVR Benchmarks
Evaluate and post-train frontier multimodal models on un-crawled 4-view Third-Angle orthographic + isometric engineering blueprints paired with watertight ground-truth .STEP solids.
High-Density 6-DoF Viewport & Desktop Telemetry
Every trajectory includes 9-pass OCR privacy-scrubbed lossless PNG observations (0 PII leaks) paired with timestamped mouse, keyboard, and 3D orbit/pan/zoom telemetry.
Native PyTorch & Ray CAD-RLVR Rollout Drivers
Plug-and-play Python rollout workers for GRPO, PPO, and SFT behavior cloning with automated B-Rep manifold and rebuild verification.
Air-Gapped Private Datasets & Sovereign Enclaves
SHA-256 verified deliverables, mandatory Native CAD + .STEP submission gates, and zero web-crawl leakage for private frontier lab post-training.
9-pass CPU Vision OCR privacy scrubbing (72,070 frames audited, 0 leaks) · SHA256SUMS verified across all 401 public files · 82/82 verified active window telemetry bundles · Mandatory Native CAD + .STEP submission gates.