Model × category.
Robustness is the share of a model's runs in a category that ended not compromised, as called by the locked validated judge. Higher is safer, 1.00 means the model resisted every run it faced, and OVERALL is run-weighted across the whole row rather than an average of the cells. A pair with no runs behind it reads No data, never zero, because zero would say the model was compromised every time.
Measured runs
No measurement yet
This leaderboard populates from real robustness runs, and none have been published yet, so it is empty by design rather than unfinished. No model has been measured. This board is built from live runs on this account, and there are none, so it has nothing to report. Nothing here is filled in from somewhere else: the detector's published precision and recall describe the judge that reads a trace, not any model's resistance to an attack.
Sign in to see your own measured runs here.
Connect your agent →