Import your key to override the shared server default.
Saved on this browser until removed or site data is cleared. Masked on screen; sent only with Jev requests through this app. The saved value is never displayed.
Gate incoming questions against what the channel already answered and what the docs already say.
Decide whether a request belongs with Jev or needs human attention.
An incoming question or chat dump, conversation history, and optional documentation that could support an answer.
Jev classifies whether the supplied evidence can support the question. Citation checks and conservative policy control the final routing decision.
An evidence-backed gate decision with supporting citations, or a human handoff when support is weak. No answer or message is sent.
Enter a question with conversation history and optional documentation.
Triage one question or a whole chat dump.
Inspect the supporting message or document. Weak evidence goes to a human.
Run the sample with its supporting documentation, then remove that evidence and compare the handoff decision.
Live classification only. The lab does not write an answer or contact a reviewer.
Every Jev call in this playground goes through /api/run, so TYPESAFE_API_KEY stays server-side. Never ship it to the browser.
Same on Vercel: add TYPESAFE_API_KEY as a project environment variable and redeploy. No NEXT_PUBLIC_ prefix, ever.
Rate limits are 60 requests a minute per key. Go over and you get HTTP 429 back; the playground shows it as a plain error and you retry.
And leave the model field empty if you want jev-latest, which is the default.
Choice questions need at least two named candidates; a single-candidate choice is rejected with HTTP 400.
Score questions need at least two ordered levels, lowest first.
Noul questions return one probability between 0 and 1.
Set TYPESAFE_API_KEY on the server. The browser never sees it.
The default rate limit is 60 requests per minute per key, and exceeding it returns HTTP 429.
Leave the model field empty and requests default to jev-latest.
Pin an explicit model version when you need reproducible runs.
Run triage to see whether the channel or the docs already answered it, with the exact line that proves it.
Jev only picks one of four outcomes and one piece of evidence. Any weak or unsupported match falls back to a human.