Workflow guide
When not to use Jev
Keep Jev on System One judgments. Do not send it work that needs generation, tools with arguments, or slow reasoning.
Step 1
No prose, no tool arguments
Jev answers typed questions. It does not write emails, patches, or function-call arguments. Pydantic-AI's TypeSafe integration even raises ToolCallProposed when a tool needs arguments.
Step 2
Split fat questions
Analyze this and decide the best course of action is a System 2 prompt. Split it into independent Noul/Choice/Score questions and combine them in code.
Step 3
Do not copy unofficial mode fields
Some community pages show a mode field such as compaction or router. Official TypeSafe requests use model, state, and questions. This hub rejects mode on purpose.
Step 4
Local Jev-like models are not Jev
jevlike and similar experiments share a type signature, not weights, calibration, or the hosted API. Label them as local experiments in any catalog you keep.
Related projects
- jevlike — Independent starter model with a Jev-like input/output shape: text plus options in, probabilities out. Not TypeSafe Jev.
- system-one-adapter-python — Official TypeSafe adapter that speaks the same system_one interface, backed by OpenAI or Anthropic instead of Jev.