Jev & System One Models — a practitioner's guide with sources

Jev is TypeSafe AI's "System One" model: it answers typed questions about text — yes/no, one of N, a rating — with probabilities, and generates nothing. This guide is what one team learned about using it well in the weeks after its launch (2026-09-15): distilled rules, the measured numbers behind them, three experiments that mostly came out negative, and a full citation index.

Scope of validity: jev-1.13.0, documentation fetched 2026-09-19. Re-check anything here after 2026-10-19 — the product is in early access and its docs move.

Evidence tags, on every non-obvious claim: [V] vendor-stated · [M] independently measured · [S] speculative or reverse-engineered · [A] the author's judgement · [M?] reported but not confirmed against the primary source.

How to read it: go to Chapter 2, find your task, load only the chapters it names. Chapters 20-23 are the sources; Chapter 23 lists the claims that did not survive checking.

What is not here. The private edition mirrors all 109 pages of docs.typesafe.ai and the typesafe-ai repositories; that is the vendor's documentation and is not redistributed — read it at https://docs.typesafe.ai. Two chapters about the author's own machine are omitted, and so is the private group discussion that seeded the research (see Chapter 20's note).

Compiled by Yotam Fromm with Claude. Quoted or summarised documentation remains © TypeSafe AI; cited measurements belong to the people who made them.