run.veric.dev
AI vertical · DP preview
Visualization preview — geometry real, lattice content M3 mock.
AI Provenance · visualizations

The order on AI-pipeline artefacts.

Two views over the same substrate. The lattice plots the partial order on AI-pipeline artefacts — raw corpora at the top, regulator-facing deliverables (Annex IV pack, DSAR attestation, FTC §5 disclosure) at the bottom. The Galois connection captures the abstraction map α from a concrete training row to the regulation region that binds it, plus its right adjoint γ.

Same geometry as the generic substrate views on /explore; the difference is the tag glossary, the colour palette, and the edge labels. The substrate is constant, the deliverable is regime-specific — see the substrate thesis.

Lattice — AI-pipeline artefacts

Hasse diagram · 12 nodes · 14 edges

Each node is an artefact in an AI training pipeline; stroke colour tracks the dominant regulation family (indigo for AI Act, sky for GDPR, amber for US state laws, rose for FTC, violet for Copyright/DSM, emerald for sectoral). Edge labels carry `ml:provenanceTags`-style annotations. The red-dashed edge is a forbidden-flow rule from Agent F's Croissant manifest 09 — `no pii reaches training-input`.

AI-provenance lattice · 12 artefacts × 14 flowshover a node · click to pin · red-dashed = forbidden-flow rule
declared-indeclared-indeclared-inml:recordSetml:Field/embedforbids: pii→training-inputsubject-ofsubject-oftrainsfine-tunesserved-atAnnex IV §2(d)GDPR Art. 17FTC §5 + AB 2013pii-corpuslicensed-imagesscraped-textcroissant-manifesttraining-corpusderived-embeddingdsar-indexmodel-checkpointdeployment-logannex-iv-packdsar-attestationftc-disclosure

Galois connection — α: row → region

adjunction · 5 regions · 24 rows

Top strip: five regulation regions (AI Act training-input, GDPR personal-data, Copyright licensed, FTC disclosure, Sectoral type-coherence). Bottom strip: 24 training rows tagged with `ml:provenanceTags`. α maps a row to the region that binds it; γ shadows back which rows a region covers. Toggle “break the connection” to force a `pii` row into the AI-Act training-input region — the wires turn red and the adjoint inset flips to VIOLATED, the visual signature of a T6 forbidden-flow refutation.

AI-provenance Galois · α: row → region · γ: region → rows
ABSTRACT · regulation regions α(c)AI Act · training-inputGDPR · personal-dataCopyright · licensedFTC · disclosureSectoral · ISO 42001α(c) ⊑ a ⟺ c ⊑ γ(a)CONCRETE · training rows c (ml:provenanceTags)r01 · tag=pii · users.csv col=emailr02 · tag=eu-personal-data · users.csv col=address (DE)r03 · tag=gdpr-erased · users.csv subject=4192813 (DSAR cleared)r04 · tag=dsar-residual · embeddings.pq subject=4192813 (NOT cleared)r05 · tag=eu-personal-data · support_tickets.json col=bodyr06 · tag=pii · logs.jsonl col=ip_addressr07 · tag=gdpr-erased · users.csv subject=8810024 (DSAR cleared)r08 · tag=synthetic · synth-augment-2026-04.jsonlr09 · tag=licensed-cc0 · openimages-cc0.tar shard=003r10 · tag=licensed-cc0 · wikitext-cc-by-sa-4.0r11 · tag=scraped-no-licence · commoncrawl-2024-43 *.htmlr12 · tag=synthetic · rl-self-play-trace-2026-q2r13 · tag=scraped-no-licence · github-noncomm-2025 *.pyr14 · tag=pii · minor-pii-fine-tune.jsonlr15 · tag=embedding-shape-mismatch · embeddings-512d cast→1024dr16 · tag=embedding-shape-mismatch · fp16 ↦ fp32 silent widenr17 · tag=synthetic · data-aug-pipeline-v2r18 · tag=licensed-cc0 · wikidata-cc0 dump 2026-04r19 · tag=embedding-shape-mismatch · adapter-projection-mismatchr20 · tag=scraped-no-licence · twitter-archive-2024 *.jsonr21 · tag=synthetic · self-play-rl 2026-q2r22 · tag=embedding-shape-mismatch · vector-store schema driftr23 · tag=dsar-residual · embeddings.pq subject=8810024 (residual)r24 · tag=scraped-no-licence · stackoverflow-archive-2025 (CC-BY-SA mismatch)α ⊣ γ5 rows · ✓
ml:provenanceTags palette
  • pii
  • eu-personal-data
  • gdpr-erased
  • dsar-residual
  • licensed-cc0
  • scraped-no-licence
  • synthetic
  • embedding-shape-mismatch
Active region · γ(gdpr)

Art. 6/9: any row whose subject did not consent or whose DSAR cleared must NOT land here.

click rows to multi-select · shift/cmd-click to add · hover a region to pin γ-shadow · 5 rows preselected
What a real verifier would do differently

These are illustrative re-skins. The lattice nodes and edges are hand-curated from the surface inventory rather than extracted from a real Croissant manifest, and the Galois tag-to-region mapping is a static lookup. A real verifier would build the lattice from the IR emitted by `silver-croissant-v2::parse`, type-check `ml:provenanceTags` against the canonical tag glossary at content/ai-tag-glossary.json, and emit a counterexample trace when α(c) ⊑ a fails. See /ai/playground for the M1 client-side approximation.