Product · October 22, 2026
A complex request is a list, not a longer prompt
Fan-out keeps each part of a hard request on its own contract, with its own model, and shows the list.
Most gateways hide a chain of models behind one answer. Separator keeps the chain visible. A request such as “close the month, write the summary, and check the contract” is three types: a workflow, a document, and a judgment.
Each part is a small request. ops-1 can reconcile. write-1 can draft. reason-1 can read a clause. They do not share a context they were not given. The parent decision lists the parts in order and records which model took which part.
Fan-out is still one API call from the caller’s side. The extra calls happen inside Separator, under the allow-list the caller sent. If a needed lane is not allowed, that part is refused instead of being quietly rerouted.
The latency we design for is the decision, not the model’s full generation. Classification and the split are the millisecond path. Generation time belongs to the model on the lane.