That’s so. For example, at Oxford and with CIP I proposed Weval.org. It lets civic organisations actually deploying these models on the ground — or even our household — you can say our lobster, this Kami: its replies to my dad on these 100 turns were bad — how should they be better? And those few turns were good — what counts as good? Give a rough description and that’s enough. Weval.org is full of that kind of thing: records of interaction or dialogue, and you say this caused a bad outcome and here’s how it could be better.