Encode the state
The state side sees the current context plus the routing question. It produces one state embedding for this hop; it does not generate a response.
RUNTIME ARCHITECTURE / 01 NS / ROUTER
The model chooses.
The router controls.
The Router turns runtime context into one ranked, legal next step. The graph owns the boundaries; the model supplies the signal.
FORWARD PASS / ROUTERWORKED EXAMPLE · ILLUSTRATIVE
01 / MISSION
“I need a copy of my invoice.”
LEGAL TARGETSBillingTechnicalFallback
02 / STATE ENCODERRepresent this request
The current hop becomes one embedding.
03 / ACTION ENCODER + SCORECompare allowed actions
Similarity scores rank only the graph's legal targets. Bar lengths are illustrative.
04 / SOFTMAX + FSM GATECommit one legal edge
RANK 01Billing
↓GRAPH EDGEApproved → BillingHelp
The runtime checks the selected edge before execution.
STATE × ACTION SIMILARITY → RANKING → GRAPH CHECKONE SELECTED TRANSITION
ROUTER / FORWARD PATH
One job.
A clear boundary.
Routing is a classification decision with an explicit action space.
Score closed choices
Each legal target is an action embedding. A similarity score compares the state with every candidate, then softmax turns the scores into a ranked probability distribution.
Gate the transition
The runtime consumes rank one and checks it against the graph. The scorer proposes; the FSM decides whether the edge can execute.
WHY THIS SPLIT MATTERS
Let the router
stay small.
Keep extraction, routing, and reasoning as separate observable decisions. Each one can be measured, tuned, and rejected without hiding the graph’s intent inside a single prompt.
Next: Schema extraction →