Code owns the workflow; Jev supplies typed judgments
Your mission is AI-powered software — not another chat agent. Jev (System One) understands language, but it returns decisions and probabilities your code can combine.
One idea
LLM agents choose the next step. Useful when a human watches; every loop can go off the rails.
AI-powered software keeps control flow, rules, and side effects in code. The model appears only where you need programmable common sense over unstructured data — as atomic, typed questions.
Primary: How to build with TypeSafe · System One
Three primitives
- Choice — one option from a fixed unordered set →
choice+ distribution +confidence - Score — position on ordered levels you define →
score+ distribution +confidence - Noul — yes/no →
noulin [0, 1] (near 0.5 means uncertain, not "medium intensity")
You send state + a map of questions; answers come back under the same IDs. Independent questions in one request run in parallel. Question IDs are for your code — put the full meaning in instructions.
Retrieval practice
Q1. Route a support ticket to billing, orders, or account. Which primitive?
Q2. "Does this PR touch auth or payments?" for a merge gate. Which primitive?
Q3. Best architecture for a CI failure fixer?
Primary source (read next)
Read How to build with TypeSafe — especially "Three software architectures" and "Decompose the questions" (~15 min). Then ask me anything unclear.
Ask your teacher
In chat: "map my PR triage onto Choice/Noul/Score" or "is this judgment too broad to ask in one question?"