New
NewHow it uses Jev
The tool-call gate sends the current task, step, proposed tool name and arguments, and explicit restrictions to Jev. Three Noul questions assess task fit, missing information and restriction violations; an optional Score rates impact if wrong.
What Jev decides
Tool-call assessmentExample answers · not a recorded Jev response · Source ↗
Question 1 · matches task
YOUR APP
INSTRUCTION
Does this tool call serve the current task or step?
STATE
task, optional step, call.tool and serialized call.arguments, restrictions, plus a notice that argument text is untrusted.
JEV · NOUL
YesNo
Question 2 · missing information
YOUR APP
INSTRUCTION
Does this call need information absent from task or arguments?
STATE
task, optional step, call.tool and serialized call.arguments, restrictions, plus a notice that argument text is untrusted.
JEV · NOUL
YesNo
Question 3 · violates restriction
YOUR APP
INSTRUCTION
Does this call contradict an explicit restriction?
STATE
task, optional step, call.tool and serialized call.arguments, restrictions, plus a notice that argument text is untrusted.
JEV · NOUL
YesNo
App workflow
- Capture a proposed call
Serialize the agent task, current step, exact tool call and restrictions.
- Ask Jev
Batch three call-specific Noul questions, optionally adding an impact Score.
- Apply gate thresholds
Use the typed probabilities and configured thresholds to proceed, request missing information, or block a contradictory action.
Why it is interesting
It converts a broad agent safety judgment into several independent, call-specific typed checks, with explicit thresholds and no need for generated approval prose.