Lineage & Provenance
Every AI Product version must declare lineage and provenance sufficient to explain its constitution, AI Product Type (APT), and AI Product qualification.
Lineage answers where dependencies come from and how they relate. Provenance records the evidence needed to understand how the current Product version was constituted.
Provenance Requirements
AIPROD should declare, where applicable:
- AI Primitives — learned models, reasoners, planners, intelligent search, optimizers, or hybrid intelligence engines;
- Model Products — governed model packages/services used as Product dependencies;
- Data Products / Data Sources — governed information used for training, inference, reasoning, grounding, or evaluation;
- Knowledge / Rules — knowledge bases, ontologies, rulesets, policies, prompts, or other semantic artifacts;
- Tools / Services — material callable services or execution tools;
- Training / Adaptation Data — where the Product constitution includes learned models or adaptation;
- Contributors — accountable teams or organizations involved in creation;
- Creation / Release Provenance — version and release events.
No single dependency class is mandatory for every AI Product except that AI Product Qualification must identify at least one qualifying direct or transitive AI Primitive path.
Lineage Requirements
An AI Product must document:
- Versioned Dependencies — exact upstream versions or compatible ranges as governed by AIPS;
- Upstream AI Products — material intelligence dependencies used to determine Derived APT;
- Data Product Dependencies — governed data/information inputs;
- Model Product / AI Primitive Dependencies — direct intelligence primitives where applicable;
- Version History — changes across releases, including model/rule/knowledge/dependency changes as applicable;
- Downstream Impact — known dependent Products or registered consumers where available;
- Qualification Paths — the dependency path(s) used as AI Qualification Evidence.
Relationship to APT
Lineage is the evidence base for AI Product Type (APT).
- Primary — no material upstream AI Product dependency; intelligence is constituted directly from AI Primitives plus other governed dependencies.
- Derived — the Product's distinct intelligence proposition materially depends on one or more upstream AI Products.
A model dependency alone does not establish Derived APT.
Qualification Provenance
AIPROD should be able to answer:
Why does this exact Product version currently qualify as an AI Product?
Example evidence path:
TransactionRiskAIP@7.2
→ FraudAssessmentAIP@4.1
→ FraudModelProduct@3.0
If FraudAssessmentAIP@5.0 removes the primitive path and becomes a deterministic non-AI Product, TransactionRiskAIP@7.2 remains historically unchanged while a new downstream version adopting 5.0 must be re-evaluated.
Machine-readable Requirements
Lineage should be:
- machine-readable;
- version-specific;
- linkable across AIPROD, AIPDS, DPP, and enterprise registries;
- suitable for knowledge-graph traversal and downstream impact analysis;
- explicit enough to distinguish AI Product, Model Product, Data Product, knowledge/rule, tool/service, and AI Primitive dependency kinds.
Example
AI Product: Fraud Detection Intelligence v2.1
- APT: Primary
- AI Primitive:
urn:model:fraud-gbt:4.0 - Data Products:
urn:dp:payment-lifecycle:5.2,urn:dp:fraud-history:3.1 - Rules / Policies:
urn:policy:fraud-thresholds:2.0 - Qualification Path: Fraud Detection AIP → Fraud GBT Model Product
- Version History: v2.0 → v2.1 changed model and threshold policy; Product Family qualification remained valid
- Downstream Impact: used by registered Transaction Risk AI Product versions
Summary
- Provenance describes the Product's constituting origins and evidence.
- Lineage describes versioned dependency relationships and impact paths.
- APT and AI Product qualification must be traceable to this graph.
Principle: AI lineage must explain both how the Product was built and why it is still an AI Product.