§ Roster
Runs.
Every model · framework combination we put through the dataset, ranked
by attack-success rate.
Observed ASR = AS / (AS + BLK). Excluded: not-triggered, no-attack-evidence, inconclusive, infra-issue, pending-judge.
★ Champion / Worst badges computed over all 14 paper runs — they don't shift with filters.
| Rank |
Run |
Model |
Framework |
Denom |
ASR ↓ |
Blocked |
Benign correct |
UI |
Open |
—