typesafe-jev-dojo

# typesafe-jev-dojo 🥋

[](https://github.com/lafollett-labs/typesafe-jev-dojo/actions/workflows/ci.yml)
[](./LICENSE)
[](.nvmrc)
[](https://www.typescriptlang.org/)

A hands-on dojo for **Jev** — TypeSafe AI's "System One" decision model: fast, **typed**
decisions (yes/no, pick-one, score-on-a-rubric), *not* an LLM. This repo exists to answer
what the hype videos never do — **how accurate is it, and how fast, on real inputs** — and
to show it off with a live, graphical, **real-time** demo (nothing pre-recorded).

> **Advisory only.** Jev returns calibrated probabilities and confidence; every scene here
> either shows the confidence or escalates low-confidence cases rather than blindly acting.



## Quick start

```bash
nvm use            # Node 26 (see .nvmrc)
npm install
npm run demo       # http://localhost:5178
```

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

## The scenes

A Node server drives six real-time canvas scenes over a single WebSocket. The server holds
the keys and makes every decision call; the browser is a pure renderer.

| Scene | What it shows |
| - | - |
| **Router** | An **endless, self-generating stream** of realistic tasks (or **type your own** and route it live, even while paused). Jev answers **3 typed questions** each — *which model tier* (choice) · *too ambiguous to act on?* (noul) · *operational blast radius* (score) — and routes each down a neon lane. Tiers climb `tool → haiku → sonnet → opus → **fable**`, matching Anthropic's own ladder (Fable 5.1 is the frontier, *above* Opus 5). The key idea follows Anthropic's routing rule: **serious work defaults to Opu