Data Integrity, 9-Pass OCR Privacy Scrubbing & Deterministic B-Rep Verification
How SpacetimeML guarantees zero-PII multimodal trajectories, SHA-256 verified CAD deliverables, mandatory Native + .STEP submission gates, and uncheatable ISO-10303-21 geometric RLVR rewards.
9-Pass OCR Verified
72,070 raw frames audited; 34,551 lossless privacy-masked PNGs retained.
SHA256SUMS Audited
15.48 GB public release with 100% cryptographic hash integrity.
Telemetry Attested
Dynamic foreground window tracking (app_context_updated_at) verified.
Air-Gapped RLVR
Isolated chmod 700 staging and single-tenant VPC rollout clusters.
Six-Stage Curation & Geometric Attestation
Every trajectory, viewport frame, and 3D solid in spacetime-ml/cad-cua passes strict automated verification gates before release.
9-Pass CPU Vision OCR Privacy Scrubbing
72,070 raw frames inspected · 0 PII leaks
Every raw CAD recording frame undergoes a 9-pass accurate CPU Vision OCR and region-masking pipeline to scrub usernames, file paths, title-block author strings, and OS notification banners—yielding 34,551 retained privacy-masked lossless PNG observations with 0 OCR errors.
SHA-256 Cryptographic Manifest Verification
401 public files · 15.48 GB verified
Every JSONL shard, episode observation archive, native CAD deliverable zip, ISO-128 blueprint pair, H.264 preview video, and storyboard PDF in spacetime-ml/cad-cua is hashed and verified against the canonical SHA256SUMS manifest with zero corrupt or truncated blobs.
Mandatory Native CAD + .STEP Submission Gates
215 deliverables · 100% episode pairing
Our upstream ingestion gate requires both a native parametric CAD file (.sldprt, .sldasm, .prt, .asm, .fcstd, .ipt, .dwg, .f3d) and an exported ISO-10303-21 .STEP/.STP B-Rep solid before marking a submission complete—producing 48 model archives with 0 orphan zips.
Dynamic Foreground Window Telemetry Verification
82/82 bundles verified · 0 stale sessions
Native C (PollForegroundAppContext), Python desktop, and WebRTC recorders continuously stamp app_context_updated_at across 260,203 desktop telemetry events. All 82/82 upstream trajectory bundles are verified with active CAD window provenance and pre-recording screen-share banner suppression.
Strict Blueprint vs. Deliverable Separation
82/82 catalog manifests sanitized
Automated ingestion pipelines (separate_blueprint_and_deliverables) isolate 2D ISO-128 reference engineering drawings into prompt_blueprint_files (48 PNG/PDF pairs) and enforce 0/82 2D drawing contamination inside 3D mechanical deliverable archives.
Air-Gapped Private RLVR Staging & Enclaves
Zero web-crawl leakage for private benchmarks
Private held-out CAD benchmarks, proprietary customer assemblies, and unreleased evaluation suites reside in chmod 700 air-gapped workspaces and single-tenant VPC enclaves—guaranteeing zero contamination in public pre-training corpora.
Deterministic .STEP Verification vs. Vulnerable VLM Judges
Why 3D engineering post-training requires mathematical ISO-10303-21 B-Rep verification rather than 2D screenshot heuristics.
Naive 2D Screenshot / VLM Judges (Vulnerable)
- ✕Silhouette Spoofing: Unhollowed solid blocks or open surface shells pass 2D rendered image similarity checks while failing physical manufacturability.
- ✕Occluded Internal Feature Blindness: VLM judges cannot verify internal counterbores, blind hole depths, or wall thickness tolerances hidden inside a housing.
- ✕Unverified Desktop Telemetry: Raw screen recordings without dynamic window polling include non-CAD browser tabs, PII leaks, and orphan deliverables.
- ✕Scale & Unit Ambiguity: A part modeled at 10x scale or with swapped principal axes appears identical in an auto-fit isometric viewport render.
SpacetimeML CAD-CUA & verify_step.py Pipeline
- Exact B-Rep Manifold Auditing: Parses
MANIFOLD_SOLID_BREPandCLOSED_SHELLentities, requiringOPEN_SHELL == 0and Euler-Poincaré graph fidelity (20% weight). - Analytical Surface & Bore Spectrum (20%): Verifies exact
CYLINDRICAL_SURFACE,CONICAL_SURFACE, andTOROIDAL_SURFACEcounts and radius spectra. - Principal-Axis BBox IoU (35%) & Volume/Area (25%): Enforces exact millimeter-scale dimensional conservation and closed-shell volume against ground truth (
R >= 0.85to pass). - 9-Pass OCR Scrubbed & SHA-256 Attested: 72,070 raw frames audited down to 34,551 zero-PII lossless PNGs across 401 cryptographically verified files.
Dataset Curation & Enterprise Enclave Standards
Operational controls governing our public Hugging Face releases and private frontier AI lab deployments.
Deterministic Out-of-Process B-Rep Verification (verify_step.py)
Rewards are computed directly from ISO-10303-21 .STEP topological graphs (CLOSED_SHELL, ADVANCED_FACE, EDGE_CURVE) outside the agent policy loop.
Zero-Orphan Episode-to-Deliverable Pairing
All 48 CAD submission archives (models/cad-XXXXXX.zip) are 100% paired with retained episode observations and 2D ISO-128 reference blueprints.
Pre-Recording Screen-Share Banner Suppression
Contributors complete mandatory pre-flight UI checks hiding browser/OS screen-share overlays prior to starting the CAD telemetry recorder.
Multi-Suite CAD Kernel Provenance
Verified native parametric deliverables spanning 6 CAD suites: SolidWorks (51), PTC Creo (9), AutoCAD (1), FreeCAD (1), Autodesk Fusion (1), and Autodesk Inventor (1).
Clean CC BY 4.0 Open Licensing & Provenance
Public release artifacts carry explicit CC BY 4.0 licensing, machine-readable schema.json definitions, and reproducible stats.json manifests.
Dedicated Single-Tenant Frontier Lab Clusters
Enterprise CAD-RLVR training environments run in isolated VPCs or on-premise enclaves with TLS 1.3 / AES-256-GCM encryption.
Review our data provenance & private CAD-RLVR architecture
Schedule a technical deep dive on our 9-pass OCR privacy pipeline, verify_step.py B-Rep reward engine, and air-gapped enterprise datasets.