jevlisted

# Jev: decisions, not text Jev is a **System One** model. You give it one *state* (the data) and any number of typed *questions*, and it returns typed decisions with real probability distributions — in ~325 ms (p50), for about **$0.000013**. It cannot write. Not a summary, not a line of code, not an explanation. What it does instead is answer **"which of these N things is it?"**, **"is this true?"**, and **"how much, on this scale?"** — with a calibrated confidence you can branch on, and with a **full probability distribution** no chat model hands you. It is cheap because output is free and questions to one state run in parallel — eight questions in one call cost $0.0000279, where asking them separately cost 4× that. It is **not** universally cheaper than a small chat model on a single trivial question; it wins on *structure*, on *fan-out*, and where the alternative is pasting a large corpus into a chat context. That is the whole trade: use it where the answer is a decision, the LLM where the answer is text. > Measured, not estimated. Every number here comes from `bench/run.py` against the > live API and is recorded in the ledger. Run `jevskill stats` for this machine's > own record. --- ## 0. Run this first **No install needed.** This skill ships scripts that use only the Python standard library, so they run straight out of the skill folder: ```bash python scripts/jev_query.py --state-file diff.txt --question-type noul --name breaks_api \ --instructions "Does the d