Voice agents can understand a request correctly and retrieve a stored approval record, but that record may no longer authorize the action today. RulingThread Voice inserts a governed decision boundary between an AssemblyAI Voice Agent tool call and a simulated consequential action. In the demo, the same invoice action is evaluated under two synthetic authority states. A stale predecessor returns HOLD, and the simulated state remains unchanged. The current successor returns ALLOW and releases exactly one process-local simulated effect. Repeating the approved action does not create a second effect because execution is idempotent. AssemblyAI powers the live browser voice session and agent interaction that invokes the action-checking tool. The RulingThread endpoint evaluates the current synthetic authority state and returns the decision together with receipt and state evidence. The recorded evidence includes the AssemblyAI session reference, authority revision, before-and-after state, decision, and observed outcome. This submission uses demo data only. It performs no real payment, uses no production authority source, and creates no external effect. Its central proof is simple: the action can remain the same while the governing authority state changes the outcome. A previous approval record does not become current authority merely because an agent can retrieve it.
Category tags: