evhub comments on Relaxed adversarial training for inner alignment

[evhub]

Actions are just language outputs—and since we ask for an action that describes a pseudo-input, hopefully we should be able to interpret it that way.