SecJev: Bringing Security Expertise to System One Decision Models

What it does

The authors report that SecJev-0.8B outperforms general Kev-9B by 20.51 percentage points in accuracy averaged equally across tasks. SecJev-Corpus combines predicting a security label with deciding whether an observation follows a supplied rule, across 14 tasks and eight sources. These are author-reported gains, not independent security validation; the project is not affiliated with TypeSafe.

How you can use it

Use the research to distinguish recognizing a security label from applying a rule you provide. The same observation can require different decisions under different policies; the corpus includes both kinds of questions instead of treating them as interchangeable.

Give the model the observation, the question and the meanings of the allowed answers. It returns numbers estimating how likely each yes/no answer, choice or rating is to be correct. Those estimates describe its judgment; they are not proof that a system is secure.