Red Hat finds Jev-style models don't reliably beat traditional guardrails

TL;DR

• On 2 October Red Hat published a benchmark of nine guardrails, among them TypeSafe AI's Jev, on prompt injection and content safety.
• Jev came first on content safety at 86.20%; on prompt injection an LLM judge, Qwen3.6-35B, led at 89.31%, ahead of a 200M-parameter classifier at 89.01%.
• The authors conclude decision models don't reliably beat LLM judges or trained classifiers on speed or accuracy; Red Hat will ship those classifiers in OpenShift AI 3.6.