Technical Truth

Aegis Sovereign is a Kubernetes-native platform built around a single principle: every AI model action must be traceable, auditable, and reversible. The core services — Model Registry, Compliance Engine, LLM Gateway, Audit Chain, and Policy Engine — communicate via an internal event bus. Every write operation generates an immutable audit entry. The Helm chart deploys all services into a single namespace with NetworkPolicy isolation and mTLS enforced by Istio.

# From the Aegis Sovereign release bundle
cd aegis-sovereign/helm

helm upgrade --install aegissovereign ./sovereign-gateway \
  --namespace aegissovereign --create-namespace \
  --set global.domain=sovereign.yourcompany.com \
  --set postgresql.auth.password="$(openssl rand -base64 32)" \
  --set redis.auth.password="$(openssl rand -base64 32)" \
  --set secrets.encryptionKey="$(openssl rand -hex 32)" \
  --set ingress.enabled=true \
  --set ingress.className=nginx \
  --wait --timeout 10m

The Compliance Evaluation Engine maps each of the 7 supported regulatory frameworks to a scored checklist of controls. Evaluations are triggered automatically by the CompliancePipelineAgent on every model registration, or on-demand via API. Results are scored 0–100 per framework, and failing controls are tagged with remediation steps by the RemediationSuggestionAgent. Evaluation artifacts (PDF, JSON) are stored alongside the model version and linked in the audit chain.

# Evaluate a model against all active frameworks
curl -X POST https://sovereign.yourcompany.com/api/v1/regulatory/evaluate \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "mdl_fraud_v4",
    "frameworks": ["eu_ai_act", "sr_11_7", "nist_ai_rmf"],
    "workspace_id": "ws-prod"
  }'

# Poll for result
curl https://sovereign.yourcompany.com/api/v1/regulatory/evaluations/eval_xyz \
  -H "Authorization: Bearer $TOKEN"
# → { "status": "completed", "scores": { "eu_ai_act": 91, "sr_11_7": 87, "nist_ai_rmf": 83 } }

Beyond compliance scoring, every request, agent hand-off, and model promotion passes through an eight-phase safety pipeline: (1) a Safety Policy Engine — a fast regex pre-screen followed by an ML classifier ensemble (Llama Guard, ShieldGemma, Azure AI Content Safety, OpenAI Moderation); (2) an Agentic Safety Guard that validates tool calls before execution (blocking destructive arguments like `rm -rf /` or `DROP TABLE`); (3) PII masking via Microsoft Presidio; (4) a cross-turn conversation monitor; (5) a multi-agent trust boundary; (6) Safety SLOs; (7) a red-team certification runner; and (8) a ModelSafetyRegistry clearance gate that blocks promotion until safety checks pass. Safety is a pipeline stage wired into every path — not an afterthought.

curl -X POST "https://sovereign.yourcompany.com/api/v1/llm/output-evals?workspace_id=ws-prod" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Summarize this customer record",
    "response": "...model output...",
    "checks": ["hallucination", "toxicity", "factual_consistency"]
  }'
# → { "status": "completed", "overall_score": 0.93, "passed": true }

Every state-changing operation in the platform writes an AuditEntry row with: the event type, actor, resource, diff payload, timestamp, and a SHA-256 hash of the previous entry. This creates a per-workspace, append-only, hash-chained log — altering any row breaks verification from that row forward, so it is tamper-evident. To make it tamper-proof even against the platform operator, the current chain head is periodically anchored to an external append-only ledger (immudb, Azure Confidential Ledger, or a WORM file). A third-party auditor can then obtain the anchor independently and confirm the chain was not rewritten. Export the full chain as CSV for external SIEM, legal discovery, or regulatory submission.

# Verify the chain from the latest entry back to genesis
curl "https://sovereign.yourcompany.com/api/v1/audit/entries/verify?workspace_id=ws-prod" \
  -H "Authorization: Bearer $TOKEN"
# → { "valid": true, "total": 62193, "head_hash": "a1b2..." }

# Export full audit log to CSV for legal review
curl "https://sovereign.yourcompany.com/api/v1/audit/export/csv?workspace_id=ws-prod" \
  -H "Authorization: Bearer $TOKEN" > audit-export-$(date +%Y%m%d).csv

The kill switch removes a model from all active traffic within 2 seconds. It works by updating the model's status to 'retired' in the registry and invalidating the routing cache in Redis — the LLM Gateway's next request cycle (every 500ms) picks up the change. For automated rollback: when a drift check produces critical severity (PSI ≥ 0.25), the platform reverts the model to its previous champion and fires the IncidentResponseAgent.

# Pull a model from all traffic immediately
curl -X POST https://sovereign.yourcompany.com/api/v1/models/mdl_fraud_v4/kill \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "reason": "Critical drift detected — PSI 0.31", "workspace_id": "ws-prod" }'

# Revert to previous champion
curl -X POST https://sovereign.yourcompany.com/api/v1/models/mdl_fraud_v4/revert \
  -H "Authorization: Bearer $TOKEN" \
  -d '{ "target_version": "mdl_fraud_v3" }'

Every model deployment is expressed as a versioned YAML manifest stored in your Git repository. The compliance_eval_id field is immutable — it chains the deployment record to the exact evaluation run that signed it off. OPA/Rego policy bundles are evaluated against each manifest on every merge to main. The GitOps operator watches the repo and applies manifests automatically after policy validation passes.

apiVersion: aegissovereign.io/v1
kind: ModelDeployment
metadata:
  name: fraud-detection-v4
  namespace: production
spec:
  model_id: mdl_fraud_v4
  version: "4.0.0"
  compliance_eval_id: eval_a1b2c3d4   # immutable — links to passing eval
  frameworks_passed:
    eu_ai_act: 91
    sr_11_7: 87
    nist_ai_rmf: 83
  bias_dir: 0.94
  robustness_score: 0.78
  approved_by:
    - role: legal
      user: sarah.chen@acme.com
      timestamp: "2026-04-18T14:22:00Z"