Jev (typesafe-ai/jev, via the Vercel AI Gateway) is an evaluation model:
it never writes text. It evaluates typed questions against a state and
returns choices, scores, and boolean probabilities with confidences. The
cheap chat model writes everything the teacher reads — Jev only decides.
State
The teacher's five Profile Questions answers, verbatim, as one text block.
Questions
All questions go in one call; they are evaluated in parallel against the same state.
| ID | Primitive | What it decides |
|---|---|---|
primary_need | Choice | which single category the teacher starts with — options are the seven categories 1.1 … 3.2 plus none |
worked_with_children | Noul | whether the teacher has worked with children before |
entry_depth | Score (3 levels) | 1 = first-time educator · 2 = some informal teaching · 3 = experienced but untrained |
low_resource_context | Noul | whether the context suggests limited materials or connectivity |
large_classes | Noul | whether the context suggests large or overcrowded classes |
Rules
- Always include a
none/otheroption on Choice questions. - Act on
primary_needonly when its confidence is ≥ 0.6. Below that, fall back to 1.1 Teaching & Learning Fundamentals — the research marks it the shared foundation for everyone. - 3.1 Ethics & Safeguarding is mandatory early regardless of the choice result — the research marks it non-optional.
- Code combines the answers into a decision. Decompose judgments into atomic questions and combine in code; never ask Jev to write prose.
Related: Profile Questions · Progression Rules · Topic Relevance Matrix