Experiments on TypeSafe's Jev, a "System One" decision model: you send a state and a map of typed questions, and it returns one calibrated, structured answer per question. Code owns the workflow; Jev supplies narrow judgments. This repo is a Bun workspace with one package per experiment.

Package	What it is
packages/lab	The original CLI: determinism, perturbation, and calibration checks on Jev through OpenRouter, plus a Haiku 4.5 comparison arm. Findings in its README and docs/how-jev-works.md.
packages/codenames	A Codenames spymaster and guesser built on the direct TypeSafe API: engine, offline eval against human play, and a web app where every card shows Jev's probability. The public demo: fan-out, calibration, and consistency as visible game mechanics.
packages/wavelength	A Wavelength round on the Score primitive. Code hides the target, Jev places every clue in one request, and the distance is the score.

Shared docs live in docs/: the model contract and findings, and the ideas backlog.

Setup

bun install
bun run typecheck
bun test

Per-package commands run from the root:

bun run lab -- --help           # packages/lab, needs OPENROUTER_API_KEY
bun run codenames -- board 7    # packages/codenames CLI; `bun run --cwd packages/codenames server/index.ts` serves the web app (needs TYPESAFE_API_KEY)
bun run wavelength -- deal 7    # packages/wavelength; scoring needs TYPESAFE_API_KEY in that package's .env

Add a package by creating packages/<name> with its own package.json and a tsconfig.json that extends ../../tsconfig.base.json.

uv run assets/make_header.py redraws the header above and assets/social-preview.png. It needs the Geist font and rsvg-convert.