How it started
A number that was wrong for six weeks.
Scribe began with an embarrassment. A research draft went out with an effect size that had been copied from the wrong column. Four people had read it. A very confident AI assistant had read it too, and had helpfully improved the sentence around it.
Nobody in that chain had the job of disagreeing. That's the whole gap. One model answering one question will always sound like a colleague who has already made up their mind, because that's exactly what it is.
So we stopped building a better answerer and started building a room: several specialists, each with one narrow job, each able to see and challenge the others' work, and a human who keeps the final say. The first version was ugly and slow, and it caught the mistake in eleven seconds.