Model Registry
Versioned model storage with champion/challenger A/B splits, auto-generated model cards, promotion gates, and MLflow / W&B / Hugging Face Hub integration.
Overview
The Model Registry is the central store for every AI/ML model running in your organisation. Each registration triggers the CompliancePipelineAgent which evaluates the model against your active regulatory frameworks, runs bias and adversarial robustness checks, and records a tamper-evident entry in the audit chain — all before the model is eligible for promotion to production.
Registering a Model
Register via the UI (Models → New Version), the REST API, or the TypeScript SDK. Supply the artifact URI (S3/GCS/Azure Blob), framework, and optional MLflow run ID. The CompliancePipelineAgent picks up the registration event within seconds.
1curl -X POST https://sovereign.yourcompany.com/api/v1/models \
2 -H "Authorization: Bearer $TOKEN" \
3 -H "Content-Type: application/json" \
4 -d '{
5 "name": "fraud-detection",
6 "version": "v4.0.0",
7 "framework": "sklearn",
8 "artifact_uri": "s3://ml-models/fraud/v4.0.0/model.pkl",
9 "mlflow_run_id": "a1b2c3d4e5f6",
10 "tags": { "team": "risk", "env": "staging" }
11 }'
12# → { "model_id": "mdl_fraud_v4", "status": "pending_eval", "compliance_job_id": "cj_xyz" }1import { AegisSovereignClient } from "@aegissovereign/sdk";
2
3const client = new AegisSovereignClient({
4 baseUrl: "https://sovereign.yourcompany.com",
5 apiKey: process.env.SOVEREIGN_API_KEY!,
6});
7
8const model = await client.models.register({
9 name: "fraud-detection",
10 version: "v4.0.0",
11 framework: "sklearn",
12 artifactUri: "s3://ml-models/fraud/v4.0.0/model.pkl",
13 mlflowRunId: "a1b2c3d4e5f6",
14});
15
16// Poll until compliance eval completes
17const result = await client.compliance.waitForEval(model.complianceJobId);
18console.log(result.frameworks); // { euAiAct: 91, srEleven7: 87, ... }Champion / Challenger A/B Splits
Promote a challenger model alongside the current champion and split live traffic between them. Both models must have passing compliance scores before a split is allowed. Traffic weights are updated in real time without redeployment.
1curl -X POST https://sovereign.yourcompany.com/api/v1/models/splits \
2 -H "Authorization: Bearer $TOKEN" \
3 -H "Content-Type: application/json" \
4 -d '{
5 "champion_id": "mdl_fraud_v3",
6 "challenger_id": "mdl_fraud_v4",
7 "traffic_split": { "champion": 90, "challenger": 10 },
8 "metric": "f1_score",
9 "auto_promote_threshold": 0.02
10 }'Auto-Generated Model Cards
A model card is auto-populated on registration using metadata from the artifact, any linked MLflow / W&B run, and the completed compliance evaluation. Cards are versioned alongside the model and exported as PDF or JSON for regulatory submission.
| Field | Source | Editable |
|---|---|---|
| Model name & version | Registration payload | Yes |
| Framework & artifact URI | Registration payload | Yes |
| Intended use / out-of-scope uses | UI form (required) | Yes |
| Training dataset description | MLflow run metadata | Yes |
| Bias evaluation — DIR by group | CompliancePipelineAgent | No (auto) |
| Compliance scores (all frameworks) | CompliancePipelineAgent | No (auto) |
| Adversarial robustness score | CompliancePipelineAgent | No (auto) |
| Audit chain entry hash | Audit service | No (auto) |
Promotion Gates
A model cannot be promoted to production unless all configured promotion gates pass. Gates are evaluated in order — a failure at any gate blocks promotion and creates a JIRA ticket via the IncidentResponseAgent.
| Gate | Default Threshold | Configurable |
|---|---|---|
| Compliance score (primary framework) | ≥ 75 / 100 | Yes |
| EU AI Act Annex IV (high-risk models) | All critical items documented | Override w/ reason |
| Bias — Disparate Impact Ratio | ≥ 0.80 | Yes |
| Adversarial robustness | ≥ 0.60 | Yes |
| Approval workflow | Legal + CISO sign-off | Yes |
| GitOps manifest validation (OPA) | Must pass all Rego rules | Yes |
EU AI Act panel
Every model row has a ⚖ EU AI Act action that opens a panel showing the Annex III classification badge (category, legal basis, rationale, and the fraud carve-out note where it applies), the production-readiness verdict with any blocking items, the nine-section Annex IV completeness checklist, and a one-click Download Annex IV Technical Documentation button. For high-risk models such as credit scoring (Annex III 5(b)), promotion to production returns 422 EU_AI_ACT_NOT_READY until the documentation is complete — see the [Compliance Frameworks Guide](/docs/compliance-frameworks) and the [Guided PoC](/docs/guided-poc).
MLOps Integrations
Link external training runs to registry entries for full lineage.
| Integration | What syncs | Direction |
|---|---|---|
| MLflow | Run ID, experiment, params, metrics, artifact URI | MLflow → Registry |
| Weights & Biases | Run metadata, bias metrics surfaced back to W&B | Bi-directional |
| Hugging Face Hub | Hub model card, tags; compliance eval triggered on import | Hub → Registry |
| SageMaker Model Registry | Compliance scores surfaced in SageMaker console | Bi-directional |
| Azure ML | Model assets linked to Azure ML experiment runs via REST | Azure ML → Registry |