Product documentation
System Design
A local-first, authoritative control plane designed for attributable AI code execution.
Execution Boundary
The core architectural principle is the strict separation between the Inference Path (model calls) and the Execution Path (side-effecting actions).
Control Plane
- Authoritative decision engine
- Local SQLite WAL ledger
- Policy bundle evaluator
- Signed event generator
Runtime
- Non-authoritative agent environment
- Submits structured proposals
- Executes only under valid permit
- Emits execution telemetry
Evidence & Audit
Every transition in the execution lifecycle produces a signed audit event. These events are cryptographically chained to prevent tampering and enable post-incident forensic reconstruction.
{
"event_id": "ev_01JK...",
"type": "ACTION_PERMIT_ISSUED",
"actor": "human_operator_42",
"digest": "sha256:7f83b...",
"policy_version": "v1.2.0",
"prev_hash": "sha256:a4bd..."
}Verification
Chain integrity can be verified independently of the runtime. By reconstructing the hash chain from the genesis event, auditors can prove that no events were deleted or reordered, and that every action was bound to a specific policy decision.