Agent Lab · Explainer
What is Jev?
Jev is TypeSafe’s decision model, available on OpenRouter. It doesn’t write text. You give it a situation and a typed question, and it returns an answer with probabilities. That makes it fast and predictable for small, repeated judgment calls.
Three kinds of question
- Choice: pick one of a fixed set of options, such as “port 3001, port 3002, or ask the owner.”
- Yes or no: is this claim true, such as “this command only touches a build folder.”
- Score: place something on an ordered scale you describe, such as low, medium or high risk.
Jev reads text only, including JSON. It doesn’t look at images, audio or video.
When I use it
- The same bounded decision comes up again and again, often in a scheduled workflow.
- The answer is a fixed choice, a yes or no, or a scale I can describe.
- The result drives routing, triage, ranking or escalation.
- Speed and cost per decision matter more than a written explanation.
When I don’t
- Plain code can get the answer exactly.
- The work needs open-ended research or a written explanation.
- The rules change a lot from run to run.
- The data isn’t cleared to leave my machine.
What I’ve learned
- Ask one thing at a time. Break a complex judgment into small questions, then combine the answers in code with explicit weights.
- Keep hard rules in code. A high score should never override a requirement that rules something out, and the model never approves an action that is reserved for me.
- Normalize before combining. Scales with different numbers of levels have to be put on the same footing before you add them up.
- Confidence isn’t correctness. It says how sure the model is, not whether it’s right, so test against examples with known answers.
Used in: Playbook 01, approvals and Playbook 03, email screening. TypeSafe’s own question-design guide goes deeper.