Small decisions. Interesting possibilities.Submit contentSubmit

How it uses Jev

Roughly every 200 ms the page turns what its sensors see into a JSON state with ego, lanes, road and pedestrian context and posts it to its own server, which sends one request with four questions: lane_action and speed_action as choices, hazard as a 0-3 score and pedestrian_yield as a noul. Code gates the answers before the car moves.

What Jev decides

One driving tick (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · lane action
YOUR APP
INSTRUCTION

Decide the lane maneuver for the next second, changing lanes only when the target lane exists, is clear and has no closing vehicle behind.

STATE

The state is the sensor view: ego position, speed and stopping distance, the current lane and adjacent lanes with their clearance, objects ahead, and any pedestrian in the path. Every tick asks all four questions in one request.

JEV · CHOICE
  1. keep_lane
  2. change_left
  3. change_right
Question 2 · hazard
YOUR APP
INSTRUCTION

Rate the overall collision hazard for ego right now, given everything in the state.

STATE

The state is the sensor view: ego position, speed and stopping distance, the current lane and adjacent lanes with their clearance, objects ahead, and any pedestrian in the path. Every tick asks all four questions in one request.

JEV · SCORE
03
Question 3 · pedestrian yield
YOUR APP
INSTRUCTION

Is a pedestrian on the road or about to enter ego's lane within stopping distance, so ego must stop and yield?

STATE

The state is the sensor view: ego position, speed and stopping distance, the current lane and adjacent lanes with their clearance, objects ahead, and any pedestrian in the path. Every tick asks all four questions in one request.

JEV · NOUL
YesNo

App workflow

  1. Read the sensors

    The browser simulation describes ego, the lanes, objects ahead and pedestrians as a JSON state on each decision tick.

  2. Send one request

    The page posts the state to its own server, which adds the TypeSafe credential and asks the four typed questions together.

  3. Gate the answers

    Code ignores a low-confidence lane change, lets a strong yield answer override speed, and forces at least a slowdown when the hazard score is severe.

  4. Act and repeat

    The car applies the gated action, scales the speed step by the elapsed time, and waits roughly 200 ms before the next tick.

Why it is interesting

All four judgments ride in one speculative request and the driving behaviour comes from confidence gates rather than a generated plan, so a slow or unsure Jev produces a gentler action instead of a broken simulation. Speed changes also scale with the time since the previous answer, so a faster model does not brake harder than a slow one.