AI Model Registry

Field	Value
Document ID	ASCEND-GOV-MR-001
Version	2026.04
Last Updated	April 2026
Author	Ascend Engineering Team
Publisher	OW-KAI Technologies Inc.
Classification	Enterprise Client Documentation
Compliance	EU AI Act Art. 9, NIST AI RMF MAP-1, SOC 2 CC8.1

Reading Time: 8 minutes | Skill Level: Intermediate

Overview

The AI Model Registry (FEAT-001) is the system of record for every AI/ML model deployed in your organization. Each model has a lifecycle (register → approve → wire to agents → deprecate), an immutable audit trail, and per-framework compliance flags (GDPR, SOX, HIPAA, PCI-DSS).

FEAT-001B extends the registry by linking agents to specific models, gating agent registration on model compliance status.

Model lifecycle

REGISTER ──► PENDING_REVIEW ──► PARTIALLY_APPROVED ──► APPROVED ──► (in service)
                                                                         │
                                                                         ▼
                                                                    DEPRECATED

Status	Meaning
PENDING_REVIEW	Just registered; awaiting compliance approval
PARTIALLY_APPROVED	Approved for one or more frameworks but not all four
APPROVED	Approved for GDPR + SOX + HIPAA + PCI-DSS
DEPRECATED	Soft-deleted; cannot be newly wired to agents

REST endpoints

All endpoints require admin auth. Base prefix: /api/v1/models.

Action	Endpoint
Register model	POST /api/v1/models
List models	GET /api/v1/models
Get model details	GET /api/v1/models/{model_id}
Update model metadata	PUT /api/v1/models/{model_id}
Deprecate model	DELETE /api/v1/models/{model_id}
Approve compliance frameworks	POST /api/v1/models/{model_id}/approve
Model history (audit trail)	GET /api/v1/models/{model_id}/history
Fleet summary	GET /api/v1/models/stats

Register a model

curl -X POST "https://pilot.owkai.app/api/v1/models" \
  -H "Authorization: Bearer <admin_jwt>" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "claude-3-5-sonnet-prod",
    "model_name": "Claude 3.5 Sonnet",
    "version": "2025-10-01",
    "environment": "production",
    "model_owner": "governance-team@example.com",
    "risk_level": "HIGH",
    "data_classification": "restricted",
    "business_unit": "Enterprise AI",
    "model_type": "llm",
    "framework": "anthropic-api",
    "description": "Primary LLM for production customer-facing agents",
    "contains_pii": true,
    "contains_phi": false,
    "contains_pci": false
  }'

Required fields: model_id, model_name, version, environment, model_owner.

Approve compliance frameworks

A model needs framework-specific approval before agents can be wired to it. Approve one or more in a single call:

curl -X POST "https://pilot.owkai.app/api/v1/models/claude-3-5-sonnet-prod/approve" \
  -H "Authorization: Bearer <admin_jwt>" \
  -H "Content-Type: application/json" \
  -d '{
    "frameworks": ["gdpr", "sox"],
    "approval_notes": "Reviewed DPA + vendor SOC 2 report. HIPAA/PCI-DSS pending BAA."
  }'

Valid frameworks: gdpr, sox, hipaa, pci_dss.

The backend recalculates compliance_status:

• All four approved → APPROVED
• Some approved → PARTIALLY_APPROVED

Each approval writes a MODEL_COMPLIANCE_APPROVED event to the immutable audit log.

FEAT-001B — Link an agent to a model

SDK 2.3.0

from ascend import AscendClient

client = AscendClient(api_key="owkai_...")

response = client.link_model_to_agent(
    agent_id="financial-advisor-prod",
    model_id=42,   # DeployedModel.id returned from /api/v1/models
)

The SDK method is a thin wrapper over PUT /api/registry/agents/{agent_id} with {"model_id": <int>}.

Server-side enforcement

Before setting RegisteredAgent.model_id, the backend validates:

Check	Failure response
Model exists in caller's organization	HTTP 400, "Model not found or belongs to a different organization"
Model is in APPROVED or PARTIALLY_APPROVED state	HTTP 400, "Cannot link agent to model with status '<status>'. Model must be approved or partially approved."

Successful links emit an AGENT_MODEL_LINKED event with compliance tags FEAT-001B, EU-AI-ACT-ART9, NIST-AI-RMF-MAP4, SOC2-CC8.1.

Retrieve a model's audit history

curl "https://pilot.owkai.app/api/v1/models/claude-3-5-sonnet-prod/history?limit=50" \
  -H "Authorization: Bearer <admin_jwt>"

Returns hash-chained entries: MODEL_REGISTERED, MODEL_UPDATED, MODEL_COMPLIANCE_APPROVED, MODEL_DEPRECATED — one row per lifecycle event.

Compliance mapping

Concern	Standard
Model risk classification	EU AI Act Art. 9
Documentation + map of AI system	NIST AI RMF MAP-1
Change management audit	SOC 2 CC8.1
Approval chain evidence	SOC 2 CC6.1, AU-6

Related docs

• Python SDK — link_model_to_agent
• Compliance — overall program
• Security Architecture

Enforcement at Submit Time (G-P0-02)

From SDK 2.5.1, ASCEND enforces model registry compliance at every action submit.

How it works:

1. SDK caller passes model_id in evaluate_action()
2. ASCEND checks the model against your organization's Model Registry
3. compliance_status must be approved or partially_approved to proceed
4. Non-compliant or unregistered models return HTTP 403

Compliance standards: SR-11-7, EU AI Act Art. 9/28

Response fields:

Field	Description
registry_checked	true when model_id was provided
compliance_status	approved / partially_approved / denied
enforcement	SR-11-7/EU-AI-ACT-ART9
source	payload (explicit) or agent_link (registered agent)

SDK example:

result = client.evaluate_action(
    action_type="model_inference",
    resource="ml_pipeline",
    model_id="gpt-4-production",
    wait_for_decision=False,
)
print(result.model_governance)
# {
#   "registry_checked": True,
#   "compliance_status": "approved",
#   "enforcement": "SR-11-7/EU-AI-ACT-ART9",
#   "source": "payload"
# }

---

Document Version: 2026.04 | Last Updated: April 2026