that's understandable, it's a long system card. i just spent awhile going back to the system card trying to understand why this experiment wouldn't count as "working on CoT faithfulness" and was very confused how it wouldn't if you had seen it. i agree they could've done more, though it's not clear to me that CoT monitoring is a basket worth filling with quite so many eggs anyways vs. activation monitoring and other techniques, specifically because there are so many ways (even just from pretraining) that it can be unfaithful.