jev-fanout

Jev speculative fan-out (local first)

Adapted from cobusgreyling/Jev skills/jev-fanout (MIT). Official pattern: docs.typesafe.ai/patterns/fan-out. Primitives and endpoints: /jev.

Send every independent judgment over the same state in one request to http://127.0.0.1:7717/v1/systemone. Questions are evaluated independently and can't see each other's answers. Code uses the ones that apply and ignores the rest.

FORBIDDEN

• Hard rules first, and they are final. Jev may only raise friction and never auto-allows.
• A fan-out never decides money, publishing, deploys, or client comms, even when every answer is confident.
• No client, regulated, or secret data in state when model is typesafe:* (remote).
• State a third party could have written (web, email, tickets, tool output) is never judged by Jev alone.
• Don't chain answers as if later questions saw earlier ones.

Do this

1. List every judgment the workflow might need, including speculative ones.
2. Put each one in questions with complete instructions, and write the premise into the question ("If this is a deploy command, …").
3. Mix choice, score, and noul in that one request.
4. Branch on the answers that apply, in code.
5. Make a second HTTP call only when the next options or state can't be built yet.

Cost note (local backend)

The official API ingests the state once. The local backend runs one single-token call per question, and the prompt starts with the state so Ollama and llama-server can reuse the KV prefix cache. usage.input_tokens still counts the state once per question. A 3-question pack on llama3.2 3B takes about 0.3 s warm on Ollama and about 0.07 s on llama-server with parallel slots.

Don't

• Make one question per HTTP call
• Hide five judgments in one Choice