Tool-use safety validation on safety-stress set 80-task
0Harm Rate (Call)RACG
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| RACGModel=Claude Opus 4, lambda=22026.06 | 0 | 0 | 100 | 2,265 | |
| RACGModel=Claude Sonnet 4.6, lambda=22026.06 | 0 | 0 | 100 | 2,722 | |
| RACGModel=Claude Haiku 4.5, lambda=22026.06 | 0 | 0 | 100 | 2,718 | |
| RACGModel=GPT-OSS 120B, lambda=22026.06 | 0 | 0 | 100 | 1,242 | |
| RACGModel=Nova Premier, lambda=22026.06 | 0 | 0 | 100 | 2,910 | |
| RACGModel=Nova Pro, lambda=22026.06 | 0 | 0 | 100 | 1,800 | |
| RACGModel=Nova 2 Lite, lambda=22026.06 | 0 | 0 | 100 | 3,534 | |
| CMTFModel=Claude Opus 4, Variant=causal2026.06 | 25 | 25 | 100 | 1,936 | |
| CMTFModel=Claude Sonnet 4.6, Variant=causal2026.06 | 25 | 25 | 100 | 2,329 | |
| CMTFModel=Claude Haiku 4.5, Variant=causal2026.06 | 25 | 25 | 100 | 2,326 | |
| CMTFModel=GPT-OSS 120B, Variant=causal2026.06 | 25 | 25 | 100 | 1,065 | |
| CMTFModel=Nova Premier, Variant=causal2026.06 | 25 | 25 | 100 | 2,490 | |
| CMTFModel=Nova Pro, Variant=causal2026.06 | 25 | 25 | 100 | 1,539 | |
| CMTFModel=Nova 2 Lite, Variant=causal2026.06 | 25 | 25 | 100 | 3,025 |