Skip to main content

Packaging & Interfaces for AI Products

AIPDS declares how the qualified AI Product version is operationally packaged and how consumers, systems, agents, or physical environments interact with it.

Packaging is not AI qualification. A Model Product package, for example, remains a Model Product unless the wider Product satisfies the AIPS AI Product qualification rules.


Packaging Forms​

An AI Product deployment may use one or more forms:

  1. API / Service Endpoint — REST, gRPC, GraphQL, WebSocket, AsyncAPI, or domain protocols.
  2. Container / Runtime Image — OCI/containerized execution unit.
  3. SDK / Library — installable runtime integration package.
  4. Agent Package — governed agent definition, tools, policies, memory/runtime configuration, and executable behavior.
  5. Reasoning / Rule / Knowledge Package — symbolic rules, ontology/knowledge base, reasoner configuration, and associated executable components.
  6. Embedded / Control Package — deployable package for robot, vehicle, device, PLC-adjacent, or other cyber-physical runtime.
  7. Model Package Dependency — a model package may be included as a dependency of an AI Product realization. A standalone Model Product is not automatically an AIPS AI Product.

All deployed AI Products remain registered and linked to their semantic AIPROD and Product-version identity.


Interface Classes​

Interfaces may include:

  • Input contracts — structured, unstructured, multimodal, sensor, event, state, or goal inputs.
  • Intelligence effects — scores, classifications, predictions, recommendations, generated content, plans, decisions, digital actions, or physical actions.
  • Invocation/execution contracts — batch, synchronous, asynchronous, streaming, interactive, continuous, embedded, or goal-driven.
  • Tool/action interfaces — external tools, APIs, workflows, commands, or actuator boundaries the Product may invoke.
  • Feedback interfaces — human review, correction, reward, outcome, or operational feedback where applicable.
  • Error and degraded-mode contracts — retries, fallbacks, safe-state, timeout, refusal, uncertainty, or handoff behavior.
  • Security interfaces — authentication, authorization, entitlement, policy enforcement, and audit hooks.

Machine-Readable Contracts​

Where applicable, interfaces should use machine-readable contracts such as:

  • OpenAPI / AsyncAPI;
  • JSON Schema / Avro / Protobuf;
  • event schemas;
  • tool/action schemas;
  • semantic contracts / JSON-LD;
  • physical-control or device protocol declarations.

Example — Symbolic Policy Reasoning AI Product​

  • Packaging: OCI container containing reasoner + governed rules/knowledge bundle.
  • Input: policy facts and case context.
  • Effect: conclusion + evidence trace.
  • Execution: synchronous API and batch validation.
  • Learning: none at runtime.
  • Evolution: governed rule/knowledge updates.

No trained model package is required.


Principle: Packaging describes the operational form of the Product. It must not be used to infer AI Product Family membership.

Marketplace consumption projection​

AIPDS is the authoritative source for realized consumption interfaces. The marketplace may project a consumer-safe subset such as port identity, port type, contract reference, availability, and supported effect classes.

The projection must not expose deployment secrets or internal runtime topology, and it must not invent ports that are not bound to the exact Product version.

AIPDS realized port
-> marketplace consumption projection
-> entitlement
-> governed consumption

A marketplace listing can therefore advertise that an AI Product is available through a synchronous API, event stream, agent protocol, embedded interface, or other governed port without becoming the execution layer itself.