Simplicity priors with reflective oracles — AI Alignment Forum