Architecture, trust boundaries, and validation model

Reference architecture

+----------------------+      +------------------------+
|  Customer Backend    | ---> |        Kern API        |
| policy + business    |      | auth, policy, routing  |
+----------+-----------+      +-----------+------------+
           |                              |
           | webhooks / polling           | task orchestration
           v                              v
+----------------------+      +------------------------+
| Customer Systems     | <--- | Execution + Validation |
| ops, analytics, ERP  |      | evidence + QA checks   |
+----------------------+      +------------------------+

Trust model

• • You own task intent, policy requirements, and downstream decisions.
• • Kern enforces request integrity, orchestration controls, and lifecycle state transitions.
• • Outputs are returned as structured evidence and status metadata for deterministic automation.

Dual-AI validation model

Validation combines two independent model evaluations before finalization. This reduces single-model bias and improves confidence in edge cases.

1. 1. Primary model scores completion quality against task policy.
2. 2. Secondary model performs an independent consistency review.
3. 3. Decision logic resolves agreement/conflict and returns auditable status.

Operational controls

1. 1. Create task with idempotency key.
2. 2. Receive task id and initial status.
3. 3. Handle retries with idempotency key reuse.
4. 4. Respect rate-limit signals and apply backoff with jitter.
5. 5. Consume final result and evidence package.