Request in, result out
The Desk architects every request. It reads the task before any model is called, answers from what's already solved, and otherwise takes the cheapest path that still clears the bar — then hands back a receipt. Under 50 ms of overhead.
What the Desk does
Task, language and difficulty — classified structurally, no model call. Then the best model in your tier by learned first-try odds, minus price and latency.
High-stakes answers verified; an honesty gate catches the wrong language, refusals when data exists, unsupported numbers. Fail → one model up.
Your documents when Knowledge is on; live web search when it's about now, with sources on the receipt.
Depth is added per request, exactly when the answer needs it.
One request, seven moves
Each move runs in the same call, in order — most of them without a model at all.
Magic-link in, connect a key, ask — you're running in a couple of minutes.