Question 1

Is Jev actually driving the signals, or is the controller doing the work?

Accepted Answer

The controller does the work. Jev answers a bounded citywide policy — corridor and region weights between 0.5 and 2, a pressure scale between 0.5 and 1.5 and a coarse hint — and the existing controller layer turns that into legal local decisions. Phase order, minimum and maximum green, yellow, all-red and starvation protection are enforced in code whatever the model says.

Question 2

How do I know the model was not tuned to make this one car look good?

Accepted Answer

That is the first thing the design rules out: the request never contains the ego vehicle's id, route or destination, and three independent scans — schema, observation frame and source — check for it. The baselines also replay the identical scenario from a single fingerprint over trip, traffic level, driver, seed and horizon, so Fixed and Adaptive see identical weather and only the signal logic differs.

Question 3

What happens when Jev is slow, rate-limited or simply not configured?

Accepted Answer

There is no fabricated Jev run. If the model is unconfigured, has no answer yet, or has outlived even the maximum hold of its last policy, the run stops and reports the classified reason — timeout, rate limit, upstream error or unreadable answer. Before the first provenance record the UI says it is still checking, and how much of a run was governed fresh versus held past its refresh window is recorded per run rather than shown on the label.

Question 4

Can I replay a run without hitting the API, and is an old trace trustworthy?

Accepted Answer

Replay is offline by design: every accepted policy is recorded with the simulated time it was accepted, and the benchmark command replays a trace with zero network calls. Two limits to know: traces do not bind a code or geography version, so an old trace is not a cross-version guarantee, and a trace whose fingerprint does not match — or that carries a policy live code could not have accepted — is refused before anything runs.

Question 5

What are the three columns being compared?

Accepted Answer

Fixed, signals on a fixed rotation with no sensing or adaptation, the do-nothing baseline. Adaptive, signals that react to local queue pressure and waits. Jev, a citywide policy from TypeSafe's jev through the Vercel AI Gateway that sees only citywide state.

Question 6

What is the difference between the Tourist and Local drivers?

Accepted Answer

Two ways of driving the same trip in the same city. The tourist follows the shortest route and sits through queues; the local knows the network and reroutes around trouble. Nothing about the driver changes the scenario, only how the car is driven.

Question 7

How is the API credential kept out of the browser?

Accepted Answer

The only place the service credential exists is POST /api/jev/policy. It travels in an Authorization header and never appears in the client bundle, the worker payload, browser state or logs; the browser only talks to its own origin's relay.

Jevtrafficsim

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community