AIP-18 — Model Interface Standard

Abstract#

Defines the vendor-neutral inference interface contract required for any model used within Gao AI OS.

This standard ensures:

• pluggability

• vendor neutrality

• deterministic integration

• policy enforcement compatibility

Design Principle#

Gao AI OS is model-agnostic.

Models MUST comply with runtime interface constraints.

Inference Interface Contract#

All compliant models MUST expose the following inputs.

Required Inputs

{
  "prompt": "string",
  "max_tokens": "number",
  "temperature": "number",
  "top_p": "number",
  "stop": ["string"],
  "structured_output_schema": "optional"
}

Required Outputs

{
  "text": "string",
  "finish_reason": "stop|max_tokens|error",
  "usage": {
    "prompt_tokens": "number",
    "completion_tokens": "number"
  }
}

Structured Output Mode#

If structured_output_schema is present, the model MUST:

• produce JSON-valid output

• respect schema

• not inject extraneous keys

Determinism Mode#

Runtime MAY request deterministic inference.

{
  "deterministic": true
}

Models SHOULD:

• minimize sampling variance

• support reproducible inference

Security Requirements#

Models MUST NOT:

• execute external code

• initiate network calls

• bypass runtime

All external interaction MUST go through GAR Tool Router.

Compatibility Declaration#

Each model pack MUST declare:

{
  "schema": "gao.model.pack.v1",
  "model_class": "M-16.2",
  "max_context": 8192,
  "supports_structured_output": true,
  "supports_determinism": true
}

Compliance Note#

This AIP does not mandate:

• vendor

• license

• architecture

It defines only interface behavior.