Skip to main content

AIPCH06 — Self-Describing

“Rich Capability Metadata Embedded”

What AIPCH06 is really asserting​

AIPCH06 is not asserting that:

“The AI Product has documentation or a model card.”

It is asserting that:

The AI Product embeds sufficient, structured, and interpretable metadata describing its purpose, behavior, constraints, dependencies, and usage context — such that it can be understood, evaluated, and safely used without requiring interaction with the producer.

Documentation is optional.
Self-description is mandatory.

The Essence (HDIP + AIPS Interpretation)​

An AI Product is self-describing if and only if:

  1. Its meaning, behavior, and constraints are explicitly encoded
  2. This information is available at the product boundary
  3. It is interpretable by both humans and machines

If understanding the product requires:

  • meetings with the team
  • reading external documents
  • reverse-engineering behavior

then AIPCH06 is not met, even if documentation exists.

Positive Criteria — When AIPCH06 is met​

AIPCH06 is met when all of the following are true:

1. The AI Product exposes complete capability metadata​

The product provides structured metadata covering:

  • purpose and intended use
  • decision capability and behavior
  • inputs and outputs (semantic + structural)
  • constraints (risk, policy, safety boundaries)
  • dependencies (AI Products, Data Products, tools)
  • usage expectations and limitations

This is not partial or inferred — it is explicitly declared.

2. Metadata is embedded, not external​

The metadata is:

  • part of the AI Product artifact (e.g., AIPROD / DPP)
  • accessible via APIs or registry
  • versioned with the product

It is not dependent on external documents.

3. Metadata supports independent evaluation​

A consumer can:

  • understand what the AI Product does
  • determine whether it fits their use case
  • identify risks, constraints, and limitations

Without:

  • contacting the producer
  • relying on informal explanations

4. Metadata is machine-interpretable​

The metadata:

  • is structured (not free text only)
  • can be parsed, queried, and validated
  • supports automated reasoning (e.g., PMDD, policy checks)

This enables:

agent-driven evaluation and composition

5. Metadata evolves with the product​

The metadata:

  • is updated with each version
  • reflects current behavior and constraints
  • does not become stale or disconnected

There is no drift between:

  • product behavior
  • product description

Negative Criteria — When AIPCH06 is not met​

AIPCH06 is not met if any of the following are true:

❌ Metadata is missing or incomplete​

Examples:

  • no description of behavior
  • missing input/output definitions
  • no constraints or limitations

This prevents safe usage.

❌ Metadata exists only in documents​

Examples:

  • Confluence pages
  • PDFs
  • slide decks

These are:

documentation, not self-description

❌ Metadata is unstructured or ambiguous​

Examples:

  • free-text descriptions without structure
  • unclear definitions of inputs or outputs
  • vague statements like “high accuracy”

This prevents machine interpretation.

❌ Behavior must be inferred​

Examples:

  • product must be tested to understand it
  • behavior discovered through trial and error
  • no explicit statement of decision logic or expectations

This creates risk and inconsistency.

Edge Cases (Important Guidance for Agents)​

Case 1: “Model card exists but limited scope”​

⚠️ Partial

Rationale:

  • covers some aspects (e.g., training, metrics)
  • but often misses:
    • composition
    • constraints
    • usage context

Case 2: “Rich metadata but not machine-readable”​

⚠️ Partial

Rationale:

  • useful for humans
  • not usable for automation or agents

Case 3: “AIPROD + DPP provide full structured metadata”​

✅ Met

Rationale:

  • complete, embedded, machine-interpretable
  • supports independent evaluation and automation

Case 4: “Metadata generated automatically but not curated”​

⚠️ Context-dependent

Rationale:

  • automation is good
  • but:
    • must reflect true intent and constraints
    • cannot be purely inferred

Evidence Signals an Agent Should Look For​

Authoritative evidence:​

  • AIPROD artifact with structured fields
  • DPP (Digital Product Passport) with trust and constraint metadata
  • accessible metadata via registry or API

Supporting evidence:​

  • versioned metadata aligned with product lifecycle
  • schema definitions for inputs/outputs
  • explicit dependency graph

Red flags:​

  • reliance on external documentation
  • missing constraints or limitations
  • vague or marketing-style descriptions
  • inconsistent metadata across versions

How an Agent Should Decide​

Decision rule (simplified):

If the AI Product cannot be fully understood, evaluated, and safely used based solely on its embedded, structured metadata, AIPCH06 is not met.

Why AIPCH06 Is Non-Negotiable​

Without AIPCH06:

  • discovery (AIPCH05) becomes ineffective
  • reuse (AIPCH08) becomes unsafe
  • composition (AIPCH20) becomes fragile
  • governance (AIPCH10) cannot be enforced

AIPCH06 enables:

  • independent consumption
  • safe reuse
  • automated evaluation (PMDD)
  • agent-driven ecosystems

Canonical Statement (for AIPS)​

AIPCH06 is satisfied only when an AI Product embeds complete, structured, and machine-interpretable metadata describing its purpose, behavior, inputs, outputs, constraints, and dependencies, enabling independent understanding, evaluation, and safe usage without reliance on external documentation or producer interaction.