TypeSafe Jev Dojo

A hands-on dojo for Jev, TypeSafe AI's System One decision model — fast typed decisions (yes/no, pick-one, score-on-a-rubric) rather than an LLM. It shows how accurate and fast Jev is on real inputs through a live graphical demo where nothing is pre-recorded.

No. With no API key the demo boots in SIM mode with an amber badge — decisions are generated locally so the graphics run and you can explore offline. Add a key and it flips to LIVE with a green badge, making real Jev calls.

An endless self-generating stream of realistic tasks, or your own typed input routed live. Jev answers three typed questions per task — which model tier, whether the task is too ambiguous to act on, and operational blast radius — and routes each down a neon lane from tool through haiku, sonnet and opus to fable.

Triage is the batching showcase: one support ticket gets a panel of about ten typed questions answered in a single batched Jev call, with every answer card filling at once under one readout. Queue repeats the panel over many tickets in parallel, auto-sorting rows by priority as verdicts land.

nvm use (Node 26 per .nvmrc), npm install, then npm run demo, which serves http://localhost:5178. A Node server drives the six canvas scenes over one WebSocket; the server holds the keys and the browser only renders.

The README marks the whole dojo advisory only: every scene either shows the confidence or escalates low-confidence cases rather than blindly acting. In the Router, REVIEW is a rare abstain reserved for when Jev genuinely cannot tell which tier a task needs, not a default for hard work.

LaFollett Labs publishes the dojo; the metadata lists an MIT licence, TypeScript, Node 26 and an OpenRouter topic. This entry is written from the README captured on 2026-10-15, whose excerpt ends inside the Queue scene description.