If we can accurately recognize good performance on alignment, we could elicit lots of useful alignment work from our models, even if they're playing the training game.| Planned Obsolescence
