Principle Ensembling — diversifies use of → Principles for SL-CAI (Appendix C.1)
Randomly sampling a different principle from sl-cai-principles at each revision step, instead of applying one fixed principle throughout, turns out to have almost no effect on the metric it would seem designed to improve: harmlessness PM scores for N=1,2,4,8,16 principles are essentially flat (constitutional-ai, §"3.4 Scaling Trends" / "Number of Principles in the Constitution", p. 9). ensembling-principles does not make individual revisions more harmless. Its actual payoff shows up one step downstream: drawing from more of sl-cai-principles produces more varied revised responses, and that diversity is what improves exploration once those revisions seed the RL stage. The lesson is that a design choice can fail its most obvious benchmark and still be worth keeping, because its value was never in that benchmark.