TypeSafe experiments

Small, interactive experiments with TypeSafe Jev: from understanding a support message to making decisions for a car in a 3D world.

The aim is to make the inputs, model judgments and resulting behaviour visible—including hesitation and mistakes. These are experiments with real API calls, not polished benchmarks.

Self driving car by Jev

The Fable implementation: four camera feeds, radar and blind-spot readings, remembered speed signs, and Jev's decision panel. Captured before starting a run; the decision values have not yet been populated.

The demos

Demo	What it explores	Local address
demo01 — Decision Lab	A support message becomes six independent judgments: owning team, human request, refund request, urgency, frustration and reproduction steps. Ordinary code combines the answers into an inspectable routing preview.	localhost:3001
demo02 — Jev Driving Lab	Jev chooses lane and target speed for a car on a two-way road. Random traffic, overtaking, oncoming cars, pedestrian crossings and remembered 30/50 km/h signs create changing situations. Four camera views, sensor overlays and a decision inspector show what is happening. Runs last 90, 120 or 180 seconds.	localhost:3002
demo03 — Fable implementation	A separate implementation of the same driving concept, being built with Fable to test another AI's interpretation of the brief. Currently under development and being tested. Its current runtime also calls TypeSafe Jev for driving decisions.	localhost:3003

driving-simulation.md is the consolidated driving brief: the intended experience and corrections, without prescribing an identical interface or implementation.

In demo01, routing is a preview: no real support ticket is sent and no refund is issued. In the driving demos, the camera views are rendered for the audience; Jev receives structured simulated sensor observations rather than camera images. Simulation and vehicle mechanics run in code. The implementations are experiments, not real autonomous-driving systems.

Run locally

You need Docker with Docker Compose (or a compatible Podman setup) and a TypeSafe API key. Node and application dependencies run inside containers; no host npm installation is needed.

From the repository root, create your local configuration without overwriting an existing one:

cp -n .env.example .env

Edit .env and set TYPESAFE_API_KEY to your own key. The existing alternative name TYPESAFE_API is also supported; TYPESAFE_API_KEY takes precedence. TYPESAFE_MODEL defaults to jev-latest.

Start whichever demo you want:

# Decision Lab
docker compose -f demo01/compose.yaml up -d --build

# Jev Driving Lab
docker compose -f demo02/compose.yaml up -d --build

# Fable implementation — work in progress
docker compose -f demo03/compose.yaml up -d --build

Open the matching localhost address from the table. The demos use separate Compose projects and ports, so they can run side by side. Stop one with, for example:

docker compose -f demo02/compose.yaml down

Evaluating messages or starting a driving run makes real, billable TypeSafe requests. Driving auto-restart continues making requests until paused or disabled.

Inspect and test

The interfaces expose the questions, observations, answer distributions and request measurements. The driving demos also support session exports.

Run local application tests inside a container:

docker compose -f demo01/compose.yaml run --rm app npm test
docker compose -f demo02/compose.yaml run --rm app npm test
docker compose -f demo03/compose.yaml run --rm app npm test

Those tests exercise application logic with fixtures; they do not establish model quality. Live sample suites and probe scripts call the actual API separately. Demo03's behaviour and validation are still evolving.

Recorded observations are in demo01/VERIFICATION.md, demo02/VERIFICATION.md and demo02/artifacts/. The saved driving traces describe synthetic situations and retain unsuccessful decisions as well as successful ones. Results apply to their recorded versions, not every later change or random seed.

Credentials and sharing

• .env and local variants are ignored by Git. .env.example contains placeholders only.
• Compose injects credentials at runtime. The root .env is outside each demo's image build context, and credentials stay in the server-side API calls.
• The services bind to 127.0.0.1 for local use. Publishing the source does not turn them into authenticated public web services.
• Demo01 sends the message you submit to TypeSafe; use synthetic examples when sharing a recording or result. Driving exports contain simulated observations and model outputs.

The TypeSafe skill was used as session guidance; it is not installed or bundled in this repository.