Question 1

Is the comparison fair, or is the simulator built to flatter Jev?

Accepted Answer

It is built to make the comparison falsifiable. Rules, Jev and any adapter you load all share the same physics, sensing, controller and emergency brake, and all choose from the same surviving manoeuvres after the same three-second lookahead — so the variable is the choice, not the car handling. You can rewind the last two minutes to just before an incident, and incident timelines keep the recorded decisions. The README also separates coaching score from pass/fail, which is the right call when the thing under test is a decision policy.

Question 2

Can I try it without an API key or a local install?

Accepted Answer

Yes. The hosted browser demo runs the free Rules and starter agents with no installation and no key. To run it locally you install uv, clone, and run the server script — no frontend build step, and bundled maps plus Three.js mean the Rules driver works without external network access once Python dependencies are in. Live Jev is a separate step: copy the env example, set a TypeSafe key, restart. The key stays on the local server, with spend and request-rate guards capping calls.

Question 3

What does it actually mean to pass a challenge?

Accepted Answer

Arrival, plus a clean record. An evaluation pass requires no audited collision, no red light, no missed stop, no failed yield, and less than one second off-road. Coaching scores cover safety, road rules, smoothness and control, but the README insists a high coaching score alone is not a pass — which is the distinction that matters when the thing being tested is a decision policy rather than a driving line. The six authored challenges sit at the arena route; seeded city routes are at the bench route.

Question 4

Can I plug in my own driving policy and have it judged?

Accepted Answer

That is one of the main entry points. You write a JavaScript agent, load it, run the fixed evaluation suite, and compare safety outcomes, comfort, latency and cost — then save the evaluation, import an earlier baseline for comparison, and export versioned results. An agent guide ships a starter adapter and the evaluation contract. There is also a perception view exposing visible hazards, building occlusion, sight limits, predicted trajectories and the reason each candidate manoeuvre was rejected, which is where you find out why your policy lost an option.

Question 5

How much can a Jev drive spend before it stops?

Accepted Answer

Two caps, both on the server. JEV_FSD_BUDGET_USD sets the spend cap at $1 per run by default and JEV_FSD_RPM caps the request rate. You enable it by copying .env.example to .env, setting TYPESAFE_API_KEY and restarting; the key stays on the local server and is never sent to the browser. The project is independent and not affiliated with TypeSafe AI.

Question 6

What is in the world besides the road, and does it need a network?

Accepted Answer

Buildings, stop signs, traffic lights, kerbside parked cars, moving traffic, cyclists and people on foot, built from OpenStreetMap data under ODbL. Eight neighbourhoods across seven Canadian cities ship with the project, so the app works offline once its dependencies are installed. The sun follows each city's real latitude, street lights come on at dusk, and Settings offers Low, High and Ultra graphics presets.

Question 7

What can I do with a drive after it has happened?

Accepted Answer

Scrub it. The last two minutes of every drive are recorded, so you can jump back to just before an incident, take over manually, switch to a different driver, or force one of the other manoeuvres the AI considered and watch it play out. Press O for photo mode, which hides the HUD and lets the cinematic director frame the shot, and Save photo downloads a clean PNG.

Jev Fsd

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community