Trust and provenance
Every claim on a dataset page traces to a file.
AI buyers are asked to document their training data. SVGZO is built so that the documentation is ready before the licence is signed.
- Provenance
- Documented per asset
- Structure
- SVG-native
- Validation
- Parse and render, two renderers
- Duplicates
- 0 byte-identical in the 850,000-page infographic engine audit; other lines measured separately
- Checksum
- SHA-256 per file and per shard
- Licence
- Research, Commercial AI, Enterprise
Data origin
Assets are composed, created and processed into SVG from source files by The Doodle Desk's team and network of creators. The company holds the rights it licenses. Every figure, name and caption is a written placeholder. No real organisation, person or measurement appears.
Rights and licensing
Every delivery carries the licence tier text, a per-asset manifest, and NOTICES/ with the notices that travel with any embedded typeface.
Manifest
Each dataset ships a per-asset manifest with identifiers, checksums and version, a Croissant 1.0 manifest and a Hugging Face-style dataset card. A copy-ready paragraph for the EU AI Act general-purpose model training-content summary is included.
Processing and validation
Every file is parsed with a hardened XML parser, rendered with resvg and cross-checked against librsvg, and inspected for viewBox, dimensions, live text, raster content and empty groups. Element counts, text counts and path counts are written into the record.
Duplicate handling
Exact duplicates are detected with a structural hash over the canonicalised DOM. Near-duplicates are detected with a perceptual hash over the render at Hamming distance 6 or less. Thresholds and rates are published per dataset. The September 2026 audit over 850,000 engine-built infographic pages found no byte-identical pages and no repeated headline and subject pairs.
Versioning and integrity
Datasets carry semantic versions. Every release ships SHA256SUMS per shard and a changelog with delta counts. Enterprise licensees receive change notices.
Delivery integrity
Files are delivered from object storage over signed URLs or scoped credentials. Buyers verify against the published checksums before extraction. Raw SVG is never served to a browser from svgzo.com; previews are deterministic WebP renders.
Takedown and grievance
Anyone who believes an asset infringes their rights can write to the address on the takedown page. Claims are acknowledged within two business days, the asset is suspended from new deliveries while reviewed, and a counter-notice process is available. A public log of resolved notices will be maintained.
Certification readiness
The documentation above is intended to supply what a Fairly Trained Licensed Model submission and an EU training-content summary ask for. SVGZO holds no certification of its own, and a licensee's submission is assessed by the certifying body, not by us.