Benchmarks / Kai / HarmBench (benign prompts)
Kai · Decision harnessHarmBench (benign prompts)
The share of HarmBench’s benign prompts each model leaves unflagged: the row Decision Index v2 scores in place of the jailbreak suite.
Latest
History
One result so far. Its history fills in as later runs are posted.
| Measured | Subject | Metric | Value | Baseline | Result | Status |
|---|---|---|---|---|---|---|
| sealed board, undated | kai-1 · 0834a74f | Accuracy | 98.6% | Jev 83.6% | Win | Frozen · sealed board |
Conditions
benign prompts only, none an attackone threshold, the sealed jailbreak board’s answersa count from the board, not a research runHardware
The run records no device or host.
Source
Decision Index v2 on Kai’s page, with the board’s note
Reproduce
No public command regenerates this result. The source above is the record it is read from.
More Kai benchmarks
AG News · DAIR Emotion · Banking77 · Support triage · Email spam · Phishing · Jailbreak · Toxicity · RAG relevance · Model routing · Typed decisions · MASSIVE · AG News, validation split · Support triage, validation split · Typed decisions, validation split · Choosing among many options · Joint decisions · Release gate against Jev · Latency of one decision
Every result for HarmBench (benign prompts) on the shelf · How these are measured