Metadata & Self-Description
A qualified AI Product must be self-describing so that both humans and machines can understand what it is, why it qualifies as an AI Product, how its intelligence is constituted, and how it should be consumed and governed.
Metadata Requirements
AIPROD should expose metadata across the following categories.
1. Identity Metadata
- Product ID and version;
- AIPROD descriptor ID;
- name and description;
- owner, maintainers, and providing organization.
2. AI Qualification Metadata
- AI Product Family qualification status;
- value-bearing essence;
- distinct intelligence proposition;
- qualifying AI Primitive evidence/path;
- qualification evaluation provenance/currentness.
See AI Product Qualification & Evidence.
3. Intrinsic AI Classification Metadata
- AI Product Type (APT): Primary or Derived;
- derivation topology where applicable;
- Intelligence Capabilities (IC).
4. Contextual / Behavioural Ontology Metadata
As applicable:
- Consumption Scope (CS);
- Purpose (PU);
- Interaction / Execution Mode (MO);
- Intelligence Effect (IE);
- Agency / Autonomy (AR);
- Embodiment (EM);
- Risk Tier (RT, R0–R4).
5. Input / Output Metadata
- input/output contracts and schemas;
- uncertainty/confidence semantics where applicable;
- evidence/explanation attachments;
- action/effect contract where the Product can cause digital or physical effects.
6. Lineage Metadata
- Data Products and source dependencies;
- Model Products and AI Primitives;
- knowledge/rules/tools/services;
- upstream AI Products;
- versioned dependency paths used for APT and qualification evidence.
7. Governance Metadata
- AIPS R0–R4 risk tier;
- policies and prohibited uses;
- control obligations;
- regulatory or organizational standards where applicable.
8. Operational Metadata
- quality/evaluation claims;
- observability hooks;
- lifecycle state;
- compatibility/deprecation metadata.
Human- and Machine-readable Forms
AIPROD metadata must be available in machine-readable serializations defined by AIPS and may be projected into human-readable Product cards or marketplace descriptions.
Model cards, training-data descriptions, or system cards are conditional evidence artifacts. They are not universal substitutes for the AI Product semantic descriptor because not every AI Product is model-centric or learning-based.
Example
AI Product: Legal Reasoning Assistant
- APT: Primary
- IC: Reasoning + Generation
- AI Qualification: Current
- Primitive Evidence: Registered LLM Model Product + governed reasoning policy
- AR: Assist
- EM: Non-embodied
- RT: Assessed AIPS risk tier
- Lineage: Versioned legal knowledge Data Products, Model Product, prompt/policy artifacts
Summary
Self-description must explain not only what the Product does, but also why it is an AI Product and how that classification remains current.
Principle: AI Product metadata is both a semantic contract and evidence for ontological classification.