Agent workflows · 2026-09-25
Designing human approval with a Jev confidence threshold
A guide to connecting Choice and Score confidence to human approval and automation using illustrative thresholds.
Confidence differs from probability
Confidence describes how well a Choice or Score answer fits the state and defined options; it does not mean an action is safe. A wrong internal-report label is reversible, while wrongly closing an account is high impact. Those actions should not automate at the same threshold. interpreting a confidence value için kararın hangi alan tarafından tüketildiğini belirtin.
Noul needs a separate policy: its noul field is the probability that the condition is yes. A likely fraud signal can send a request to review rather than punish a customer automatically. Show reviewers the state, decision, and proposed action together, then record accept, reject, and correction. Sonuçları gerçek örneklerle inceleyin; istisnayı soru metni, seçenekler ve iş kuralı arasında görünür tutun.
Choose a threshold by the risk of the action
Noul needs a separate policy: its noul field is the probability that the condition is yes. A likely fraud signal can send a request to review rather than punish a customer automatically. Show reviewers the state, decision, and proposed action together, then record accept, reject, and correction. reversible and high-impact actions için kararın hangi alan tarafından tüketildiğini belirtin.
Confidence describes how well a Choice or Score answer fits the state and defined options; it does not mean an action is safe. A wrong internal-report label is reversible, while wrongly closing an account is high impact. Those actions should not automate at the same threshold. Sonuçları gerçek örneklerle inceleyin; istisnayı soru metni, seçenekler ve iş kuralı arasında görünür tutun.
| Need | Type | Field |
|---|---|---|
| Fixed option | Choice | choice, confidence |
| Ordered level | Score | score, confidence |
| Yes/no probability | Noul | noul |
An illustrative three-zone example
Confidence describes how well a Choice or Score answer fits the state and defined options; it does not mean an action is safe. A wrong internal-report label is reversible, while wrongly closing an account is high impact. Those actions should not automate at the same threshold. automation human approval and stop zones için kararın hangi alan tarafından tüketildiğini belirtin.
Noul needs a separate policy: its noul field is the probability that the condition is yes. A likely fraud signal can send a request to review rather than punish a customer automatically. Show reviewers the state, decision, and proposed action together, then record accept, reject, and correction. Sonuçları gerçek örneklerle inceleyin; istisnayı soru metni, seçenekler ve iş kuralı arasında görünür tutun.
Read a Noul result separately
Noul needs a separate policy: its noul field is the probability that the condition is yes. A likely fraud signal can send a request to review rather than punish a customer automatically. Show reviewers the state, decision, and proposed action together, then record accept, reject, and correction. direct use of a Noul probability için kararın hangi alan tarafından tüketildiğini belirtin.
Confidence describes how well a Choice or Score answer fits the state and defined options; it does not mean an action is safe. A wrong internal-report label is reversible, while wrongly closing an account is high impact. Those actions should not automate at the same threshold. Sonuçları gerçek örneklerle inceleyin; istisnayı soru metni, seçenekler ve iş kuralı arasında görünür tutun.
Illustrative threshold example
const response = await client.post("/v1/systemone", { model: "typesafe/jev", state, questions });
const answer = response.answers.route;
if (answer.confidence < 0.80) sendToHuman(answer.choice);
else routeTicket(answer.choice);Audit thresholds with your own data
Confidence describes how well a Choice or Score answer fits the state and defined options; it does not mean an action is safe. A wrong internal-report label is reversible, while wrongly closing an account is high impact. Those actions should not automate at the same threshold. measuring thresholds against past decisions için kararın hangi alan tarafından tüketildiğini belirtin.
Noul needs a separate policy: its noul field is the probability that the condition is yes. A likely fraud signal can send a request to review rather than punish a customer automatically. Show reviewers the state, decision, and proposed action together, then record accept, reject, and correction. Sonuçları gerçek örneklerle inceleyin; istisnayı soru metni, seçenekler ve iş kuralı arasında görünür tutun.
Improve the question by measuring outcomes
Noul needs a separate policy: its noul field is the probability that the condition is yes. A likely fraud signal can send a request to review rather than punish a customer automatically. Show reviewers the state, decision, and proposed action together, then record accept, reject, and correction. interpreting a confidence value için kararın hangi alan tarafından tüketildiğini belirtin.
Confidence describes how well a Choice or Score answer fits the state and defined options; it does not mean an action is safe. A wrong internal-report label is reversible, while wrongly closing an account is high impact. Those actions should not automate at the same threshold. Sonuçları gerçek örneklerle inceleyin; istisnayı soru metni, seçenekler ve iş kuralı arasında görünür tutun.
Frequently asked questions
Is 0.80 a universal threshold?
No. It is only an illustrative operating point and must be validated against error cost in your domain.
Does low confidence mean an error?
No. It means the result is uncertain and may call for more information or human review.
Does the same threshold apply to Noul?
Noul has no separate confidence; evaluate its yes probability with your own process rule.