JevBranching ≠ resolution
K is the number of choices at a refinement step. Resolution is the final grid spacing. A larger K changes the tree’s depth, not the requested precision.
A CONTINUOUS-CONTROL NOTEBOOK
Small decisions.
Continuous control.
From choosing an interval to steering a car. JevNext turns Jev’s discrete choices into numeric actions, one refinement at a time.
4environments
—recorded attempts
1–4control dimensions
Real rollouts.
Outcomes and settings included.
Outcomes and settings included.
01 / IN THE ENVIRONMENTS
Decisions, in motion.
Watch the featured recordings after prompt and controller refinement, then inspect every development attempt. Videos show simulation time; API waits are omitted. These selected examples are not estimates of general success rates.
Loading experiment records…
A CLOSER LOOK
Branching and precision.
With the revised prompt, all four K / precision settings completed the seed-7 track. An additional seed-19 track also completed. Earlier prompts and their outcomes remain in the full racing history.
02 / BEHIND THE ACTIONS
A number made
of choices.
JevNext subdivides a numeric range, asks Jev which interval to keep, and refines it until one grid cell remains.
Read the algorithmTelemetry in. Controls out.
The controller receives structured state information. Numeric outputs drive the game’s continuous action space on a finite grid. The model does not receive the rendered gameplay.
No training. Inspectable attempts.
These are inference-time control experiments with no model training. Each game’s run ledger includes recorded unsuccessful attempts, settings, timing, and available media.
03 / THE EXPERIMENT RECORD