Optimally constructing monitoring protocols with multiple monitors of varying costs and performances.

We show that penalizing certain CoT reasoning makes LLMs learn encoding schemes that generalize to unseen examples.