LinguaGuard — original wording balanced accuracy — scanner control without user request: leaderboard

Original-wording TSBench AgentDojo: 1,220 steps, 868 safe, 352 unsafe, 589 trajectories, 58 distinct user instructions. Balanced accuracy of supplied-action risk classification. Condition: scanner control without user request. Managed models three repeats; self-hosted greedy single run. Appendix E explicitly identifies Qwen3.8-27B FP8 and Sonnet5 thinking disabled. Native trajectory guards predict prefix-level risk against step labels and are not pooled with step-level guards. Scanner withholds user request. Provider defaults do not resolve unknown checkpoint variants; qualified names remain distinct. FPR/TPR and translation shifts are operating-point diagnostics, not standalone general-capability scores. Metric: Balanced accuracy. Higher is better. Source: https://arxiv.org/html/2610.22244v1

Metric: Balanced accuracy. Source: arxiv.org. 1 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=linguaguard-original-wording-balanced-accuracy-scanner-control-without-user-request · How It Works · Data refreshed daily, snapshot 2026-10-17.