Action trace from the demo: a requisition for 37 bags of cement created after policy checks pass, with the audit log written and the downstream adapter confirming the mutation.
An action trace from the public demo, with a synthetic caller.

Called for real through Bland's Custom API; an identical replay returned the same requisition instead of a duplicate.

Layman's Version

A phone agent can understand 'order 37 bags of cement for Site A', but understanding isn't permission. Someone has to check who's calling, whether they're allowed to order that, whether the order already went in, and keep a record.

This lab is that layer. It isn't a voice model. It takes what the agent extracted and decides whether it becomes a real action, needs a person's approval, or gets refused.

Technical Version

One path for every action

  1. transcript
  2. extraction
  3. caller
  4. policy
  5. action
  6. audit

A single action gateway is used by the simulator, the demo, the webhook, the evals and trace verification. Extraction is deterministic by default; optional LLM extraction is validated and can't authorize anything.

Verification

npm run eval reports 50/50 with 0 unsafe actions. That's a deterministic regression over scripted scenarios: it shows the policy engine behaves as specified, not that a model or a live deployment is reliable.

The hardened version was also called for real through Bland's Custom API against a temporary deployment with an authenticated webhook. It created a requisition in Neon, and an identical replay returned the same requisition. Those temporary resources were deleted afterwards.

The demo's policy decision view for a request that needs approval.
A request routed by policy.

Limitations

  • Operator routes use one shared secret, not user accounts.
  • Live phone calls, full conversational routing and a real ERP weren't verified.
  • The public demo isn't a customer deployment.