Skip to main content

AI Product Qualification & Evidence

AIPROD must describe why a specific Product version currently qualifies for the AI Product Family.

AI Product qualification is not a permanent label attached to a Product identity. It is a version-specific, evidence-backed ontological claim derived from the Product's current Product Intrinsic Nature (PIN) and constitution graph.

Principle: Productness and AI Product Family qualification are separate questions. A Product may be valuable, governed, and fully productized without qualifying as an AI Product.


1. Qualification Conditions​

A Product version qualifies as an AI Product only when all of the following are satisfied:

  1. Product qualification — the candidate first qualifies as a Product under BPS.
  2. Intelligence primacy — artificial intelligence is the Product's value-bearing essence.
  3. Distinct intelligence proposition — the Product exposes an identifiable governed intelligence capability of its own.
  4. Qualifying AI Primitive path — its direct or transitive constitution contains at least one qualifying Artificial Intelligence Primitive.
  5. Governed Product boundary — ownership, lifecycle, contracts, controls, versioning, observability, evaluation, and governance are defined.

See AI Product Ontology and What is an AI Product?.


2. AI Qualification Evidence (AQE)​

Each AI Product version must carry sufficient AI Qualification Evidence (AQE) to explain and validate its AI Product Family classification.

At minimum, AQE should describe:

Evidence elementMeaning
Qualification targetThe exact Product version being evaluated
Value-bearing essenceEvidence that artificial intelligence itself is what is being productized
Intelligence propositionThe distinct governed intelligence capability exposed by the Product
Qualifying AI Primitive(s)The direct or transitive intelligence-producing primitives supporting qualification
Dependency path(s)Version-specific path from Product to qualifying primitive(s)
Qualification statusCurrentness state of the qualification
Evaluation provenanceWhen and against which versioned dependencies the qualification was established

AQE is not intended to duplicate all lineage metadata. It provides the minimal evidence necessary to prove AI Product Family membership, while Lineage & Provenance carries the broader dependency and provenance graph.


3. Qualification Status​

AIPS uses qualification currentness to distinguish a historically valid classification from a currently valid one.

The conceptual statuses are:

  • CURRENT — the evidence remains valid for the declared Product version and its bound dependencies;
  • STALE — relevant upstream or constitutional information changed and the evidence may no longer reflect current reality;
  • REVALIDATION_REQUIRED — a material change requires explicit recomputation or governance review;
  • INVALID — the Product version no longer satisfies AI Product qualification.

A Product version whose qualification is INVALID must not be published or represented as a current AI Product version.

Historical versions retain their historical evidence and classification.


4. Direct and Transitive Primitive Evidence​

A Primary AI Product commonly reaches an AI Primitive directly:

A Derived AI Product may reach a qualifying primitive transitively through one or more upstream AI Products:

The transitive path is part of qualification evidence. The downstream Product does not need to own the ultimate primitive, but its current constitution must retain a valid path to one.


5. Qualification Invalidation​

Qualification must be re-evaluated when a material change can alter the Product's AI Product Family membership.

Typical triggers include:

  • removal or replacement of a qualifying AI Primitive;
  • reclassification of an upstream dependency from AI Product to another Product Family;
  • migration to a new upstream AI Product version whose qualification differs;
  • loss of the Product's distinct intelligence proposition;
  • change to the Product's value-bearing essence;
  • material restructuring of the dependency graph used to establish AI Product Type (APT).

Example:

If A v2 has no other direct or transitive qualifying AI Primitive and no longer satisfies the AI Product definition, it must not inherit A v1's AI Product Family classification merely because the logical Product identity is unchanged.


6. Downstream Impact​

A qualification-affecting change to an upstream Product does not automatically mutate every downstream Product.

Impact is evaluated version by version:

  • a downstream version still bound to an older valid upstream version remains unchanged;
  • a downstream version adopting the new upstream version must recompute its own qualification evidence;
  • an alternative AI Primitive or AI Product dependency may preserve qualification;
  • where the only qualifying path disappears, the downstream Product version may itself require Product Family reclassification.

This enables registry-driven Intrinsic Classification Impact Analysis in a mature HDIP ecosystem.


7. Qualification Is Not Self-Declaration​

The AI Product Owner may provide intent and evidence, but a mature AIPS/HDIP implementation should derive or validate qualification using authoritative registries and versioned dependency information.

AIPROD therefore records the validated semantic result and its evidence, rather than treating productFamily: AI as an unverified free-text assertion.


8. AIPS 0.1 Machine-readable Representation​

The AIPS 0.1 Core and AIPROD authoring bindings now represent qualification evidence using this structure:

hasAIQualification:
type: AIQualificationEvidence
status: CURRENT
valueBearingEssence: ARTIFICIAL_INTELLIGENCE
intelligenceProposition: transaction-risk-assessment
evaluatedAt: 2020-07-21T12:00:00Z
evaluationMethod: AIPS-AQ-001-v0.1
primitiveEvidence:
- type: AIPrimitiveEvidence
primitiveType: LEARNED_MODEL
primitiveRef: urn:model:fraud-detector:4.1.0
dependencyPath:
- urn:aiprod:transaction-risk:7.2.0
- urn:aiprod:fraud-assessment:4.1.0
- urn:model:fraud-detector:4.1.0

The normative RDF/SHACL vocabulary and informative JSON authoring bindings are published under static/spec/aips/0.1. For INVALID evidence, primitiveEvidence may be absent and invalidationReason is required; non-INVALID qualification evidence requires at least one primitive evidence path.


Summary​

  • AI Product qualification is version-specific and evidence-backed.
  • AIPROD records why the Product's current PIN qualifies it for the AI Product Family.
  • At least one current direct or transitive path to a qualifying AI Primitive is required.
  • AI Primitive presence is necessary but not sufficient: intelligence must also be the value-bearing essence and form a distinct governed proposition.
  • Qualification changes can trigger downstream graph impact analysis without rewriting historical Product versions.

Principle: AI Product is a claim that must remain provably true, not a label inherited forever.