From raw log to model weight, auditable end-to-end.
STACK Lineage tracks every byte that influenced a deployed model — raw source, cleaning step, feature, embedding, fine-tune — and emits EU AI Act Annex IV documentation on demand as signed verifiable credentials.
Regulators ask what data trained this model. Have the answer ready.
EU AI Act, NIST AI RMF, ISO 42001, and the Colorado AI Act all demand training-data provenance. Lineage produces it as a queryable, signable, auditor-ready artifact.
Source-to-Weight Graph
Every training run links to its input data sets, feature transforms, eval results, and reviewer signoffs.
Cleaning-Step Provenance
Each PII redaction, dedup, filter, and synthetic-data injection recorded as a verifiable transformation step.
Embedding Lineage
Vector store contents linked back to original document, ingest run, and license tier.
Auditor Export
Annex IV / NIST AI RMF / ISO 42001 evidence packs generated as W3C Verifiable Credentials. Signed, time-stamped, immutable.
Drift to Approved State
Detect when production model's effective training data differs from the signed baseline — silent re-training caught.
Multi-Tenant Scoping
Per-customer model lineage isolated and exportable separately. Defense and FSI deployment patterns supported.
→ Model: fraud-detector-v3 (PyTorch, fine-tuned 2026-05-10)
→ Sources: 4 training datasets, 2 feature stores, 1 synthetic gen
→ Transforms: 14 cleaning steps (PII redaction ×3, dedup ×2, filter ×9)
→ Embeddings: 847K vectors — all linked to source documents + license tier
→ Eval: 3 eval runs, 2 reviewer signoffs, 1 bias audit PASSED
→ Annex IV: EU AI Act technical documentation generated SIGNED
→ CREDENTIAL W3C Verifiable Credential emitted — auditor-signable in 1 day
From instrumentation to signed Annex IV, in four weeks
Lineage hooks into your existing ML pipeline — Airflow, Spark, HuggingFace, PyTorch. One-line instrumentation per source, not a rip-and-replace.
1. Instrument
Hooks into ingest pipeline, feature store, and training framework. One-line integration per data source and transform step.
Week 1–22. Map
Source-to-weight graph built for first production model. Cleaning steps, embeddings, and eval runs linked with cryptographic signatures.
Week 2–33. Certify
Annex IV documentation generated. W3C Verifiable Credentials signed. Auditor walkthrough completed and evidence pack delivered.
Week 3–4Questions teams ask before deploying
Straightforward answers about scope, integration, data handling, and rollout.
Do you replace Weights & Biases or MLflow?
No — we integrate with them. They track experiments. We track regulatory provenance with cryptographic chain-of-custody.
EU AI Act applicability?
High-risk system Annex IV technical documentation, foundation-model GPAI documentation, plus the conformity-assessment workflow — out of the box.
How is data captured?
Hooks into your ingest pipeline (Beacon, Airflow, Spark, Snowflake), feature store (Tecton, Feast), and training framework (HuggingFace, PyTorch, JAX). One-line instrumentation.
On-prem and air-gapped?
Yes. Verifiable credentials signed offline; transparency log can be a private Rekor instance. Air-gapped via offline-signed import.