Three ML-powered signals in one sub-200ms call — risk classification, intent relevance scoring, and prompt injection detection — gate every agent tool call before it executes in production.
Scenarios
Choose a scenario or fill in the form and hit Evaluate.
The LLM agent decides to call a tool like execute_bash or write_file.
Three typed questions — Choice, Score, Noul — classify risk, score relevance, and detect injection in a single API call.
Tunable thresholds approve safe calls, escalate ambiguous ones, and block dangerous or injected calls outright.
Agent LLM → proposes tool_call(name, args, reason)
↓
Guardrail API → POST /api/jev/evaluate
↓
Jev (System One) → 3 typed questions in 1 call (~$0.00004)
├─ Choice: safe | sensitive | destructive
├─ Score: 0.0 (unrelated) → 2.0 (direct match)
└─ Noul: P(injection) 0.0 → 1.0
↓
Decision Engine → configurable thresholds → approve | escalate | block