Compile-time provenance for AI.
Information-flow verification is the structural problem behind every regulated-data regime in 2026. Feed veric a typed pipeline IR — a Croissant training-data manifest, a dbt SQL graph, a model fine-tune script — plus a tag glossary, and it returns either a proof that no forbidden flow exists or a counterexample trace witnessing one.
Same compiler. Same proof contract. Same counterexample shape. Different tag glossary, different output template — Annex IV pack, DSAR chain-of-custody, Article 53(1)(d) summary.
Four ways into the AI-provenance vertical.
The substrate is constant. What changes per regime is the tag glossary and the deliverable template. Pick the surface that matches your role.
- T0–T9 · AI tag glossaryWhich tier proves which property?
The same ten-tier ladder, with an AI-flavored tag overlay. T6 information-flow on Croissant manifests, T7 erasure-completeness against DSARs, T1 type-coherence on embedding shapes.
Open the glossary → - 8–10 incidents · 2020–2026What goes wrong without compile-time provenance.
NYT v OpenAI. Italian DPA's ChatGPT ban. Air Canada's chatbot-as-contract. Each entry maps the failure to a tier, names the regulatory anchor, and shows what an AG-tower-driven control would have caught.
Read the archive → - Worked examples · canonical playgroundSixteen scenarios with verdicts.
PII in training corpus. gdpr-erased residual. Copyrighted text in Article 53(1)(d) summary. Embedding leakage. Each scenario ships a Croissant manifest, a pipeline fixture, and an AI-flavored SARIF verdict. Hosted by the substrate at playground.veric.dev.
Browse the examples → - Design partner previewCompliance Vault.
One source of truth for AI Act Annex IV, Article 53(1)(d) training-data summary, and DSAR-erasure attestation. Mocked end-to-end now; real customer flows live with our first design partner.
Open the preview →