A little road. A lot of decisions.

1×, 2×, 4×, 8× and 16× change road speed between decisions. API latency is real and is never divided by the speed setting. Only one request is in flight at a time.

Image mode sends a PNG. Structured scene mode sends obstacle lanes and coordinates as JSON, with no image. Jev latest and Jev 1.14 accept both input modes; the older Jev 1.13 option remains a JSON baseline. Compare models using the same input mode to compare their decision latency. A text result is not a vision accuracy result. Both modes include the selected one, two or three questions. The display stays HD; the API image can be compact or full resolution. Image mode sends no obstacle coordinates. Neither mode supplies the chosen answer label. The chosen lane becomes a left / stay / right steering command in game code. One answer controls the entire lane change. The default takes two looks: when the barrier appears and before it reaches the car. Teaching tour adds a middle look; Fast pass takes one. These distance checkpoints arrive sooner at higher speeds. The countdown is computed locally, excludes API waits, and is never sent to the model.

Classroom mode freezes both the road and game clock during every API request. It resumes from the same position when the answer arrives. Real-time mode keeps moving during requests and skips rechecks that cannot finish within the measured latency plus steering budget. It always requests a first decision for a new barrier. Neither mode changes a model’s chosen lane. When a response arrives, code checks whether the captured obstacle row still exists and the run is still active. Old responses are logged and skipped. A valid response can still be wrong and cause a collision.

Each run stops on a collision, an API error, or after 30 calls. Pause stops new calls. Calls already in flight are still recorded. The two other questions describe the scene; only the open-lane choice steers.