Partner With SpacetimeML
Scale your frontier model post-training with deterministic 3D CAD Computer-Use Agent (CUA) trajectories, ISO-128 engineering blueprints, and ISO-10303-21 B-Rep geometric RLVR verification.
401 Files
5.53 GB Public Release
SHA256SUMS verified · CC BY 4.0
62 Episodes
6 CAD Software Suites
11,792 PNGs · 237,579 events
215 Files
CAD Deliverables
115 Native + 98 .STEP + 48 Blueprints
85–90%
Frontier Model Fail Rate
Verified via verify_step.py
Tailored 3D CAD & RLVR Infrastructure
Designed for frontier AI labs, spatial capability evaluators, industrial engineering teams, and expert CAD contributors.
Frontier AI Labs
Scale multimodal Computer-Use Agents (CUA) and CAD code synthesis models with multi-suite 3D CAD trajectories, 4-view ISO-128 blueprints, and deterministic B-Rep reward verification (verify_step.py).
Key Capabilities Included:
- Private multi-suite CAD CUA episodes & native deliverables
- Deterministic ISO-10303-21 .STEP B-Rep reward engine
- PyTorch & Ray GRPO / PPO rollout workers
- 9-pass OCR privacy-scrubbed lossless viewport frames
Spatial AI & Model Evaluators
Benchmark frontier foundation models on zero-contamination ISO-128 orthographic + isometric engineering blueprints where Claude 4.5 (15.0%), GPT-5.2 (10.0%), and Gemini 3 (10.0%) fail 85–90% of tasks.
Key Capabilities Included:
- Held-out ISO-128 4-view engineering blueprint suites
- Watertight CLOSED_SHELL & Euler-Poincaré topology audits
- Analytical surface & internal bore radius spectrum scoring
- SHA-256 attested evaluation scorecards
Industrial Robotics & Manufacturing AI
Train domain-specific engineering copilots and manufacturing agents across commercial and open CAD suites (SolidWorks, PTC Creo, AutoCAD, FreeCAD, Autodesk Fusion, Autodesk Inventor) inside air-gapped VPC enclaves.
Key Capabilities Included:
- Native parametric feature-tree datasets (.sldprt, .prt, .fcstd, .ipt, .f3d)
- Assembly constraint & bill-of-materials verification
- Air-gapped single-tenant VPC or on-premise deployment
- Custom geometric tolerance & DFM reward rubrics
Senior Mechanical CAD Engineers
Contribute complex parametric parts, multi-component assemblies, and ISO-128 blueprint reconstructions across SolidWorks, PTC Creo, Fusion, and Inventor with verified native + .STEP deliverables.
Key Capabilities Included:
- Paid per-part & per-trajectory CAD bounties
- Mandatory Native CAD + .STEP submission pipeline
- Automated 9-pass OCR privacy scrubbing on all frames
- Direct impact on frontier spatial AI capabilities
Infrastructure Built for Deterministic 3D Post-Training
Why frontier multimodal and CUA research teams rely on SpacetimeML for spatial and mechanical CAD intelligence.
Live Public Baseline: spacetime-ml/cad-cua (5.53 GB)
401 SHA-256 verified files across default, models, and benchmark configs: 62 CAD episodes, 11,792 privacy-masked PNG observations, 237,579 telemetry events, 48 ISO-128 blueprint pairs, and 215 CAD deliverables (98 .STEP solids).
Deterministic B-Rep Verification (verify_step.py)
Computes exact composite reward R in [0, 1] directly from ISO-10303-21 .STEP files: 35% Principal-Axis BBox IoU, 25% Closed-Shell Volume/Area, 20% B-Rep Topological Graph & Euler-Poincaré Fidelity, and 20% Analytical Surface & Bore Radius Spectrum.
9-Pass CPU Vision OCR Privacy Scrubbing (0 Leaks)
72,070 raw desktop recording frames inspected across 9 OCR passes to scrub usernames, file paths, and OS banners—retaining 11,792 lossless PNG frames with 0 PII errors.
82/82 Verified Telemetry & Mandatory Deliverable Gates
Dynamic foreground window tracking (app_context_updated_at), strict separation of 2D prompt blueprints from 3D deliverables, and mandatory Native CAD + .STEP upload gates with 0 orphan model archives.
Ready to scale 3D CAD CUA & B-Rep RLVR?
Speak with our systems engineering team to configure private CAD trajectories and deterministic B-Rep reward harnesses.