Rules and workflows
Every one of the 15 intents, rendered as the actual decision tree the agents walk. Each outcome below was produced by executing the served autonomy handler on an embedded probe, so this tree is the agents' rules. Edit a knob and it hot-reloads into the next run.
Choose a customer question to see how your assistant handles it.
The never-auto floor (dispute_chargeback, adverse_event, and every hard legal/safety branch) is set in code and rendered locked. It cannot be relaxed from here, by design. A leaf that routes to a human because your autonomy ceiling is L0 (for example complaint) is editable: raise its ceiling to change the flow.
Detailed rules and revision history
One card per kind of message
Each card shows how the executed tree handles that kind of message, the policy clause it evaluates by name, and the writing guidance the writer reads. An edit goes through a candidate diff first, and only the reviewed bytes can be published. Rehearsing runs the real flow with every write refused; what it shows is information, and it never gates a publish.