Longer pieces for the people who have to evaluate this — architecture, evidence and limits, including the measurements that did not come out the way we hoped.
A model cannot tell what you asked from what it read on the way. How the Gateway spots an instruction nobody gave — measured one language at a time, with what it misses published next to what it catches.
Read →A simple question should not cost what a hard one costs. How the Gateway tells them apart without sending the request anywhere to find out — and what a public benchmark said when we tested the idea.
Read →