Agent Escalation

Turns uncertainty into a control signal. A support agent has proposed an action (a refund, a label, a credit); before it runs, a decision gets a confidence, and the confidence picks the path: ≥ 0.9 runs automatically, 0.6–0.9 goes to a stronger reviewer (the frontier model), < 0.6 goes to a human. Jev gives the verdict as a Choice (approve / reject) with the policy in its state, and its confidence sets the band. Against it: the same bands driven by the LLM's self-reported confidence, and reviewing every action with the frontier model. Measured: wrong actions executed (must be 0), good actions run automatically, share sent to review and to a human, and cost.